Gift personalized customization method and system based on multi-mode AIGC

Abstract

The invention relates to the technical field of artificial intelligence design, and discloses a gift personalized customization method and system based on multi-mode AIGC. The method comprises the following steps: acquiring multi-modal input information such as text description, reference images, style labels and gift purpose categories provided by a user, and performing dynamic sampling optimization on the acquisition process; performing joint coding on multi-modal input information by using a multi-modal encoder to generate multi-modal feature embedding, and analyzing a user design intention vector through an intention decoder; based on the user design intention vector, using a diffusion model to iteratively denoise to generate a gift design first draft image, and adjusting the diversity of the generation process through a condition control network; performing local detail reconstruction on the gift design first draft image by applying a content perception mechanism, and outputting an optimized gift design image; and converting the optimized gift design image into an editable design template, and integrating an augmented reality engine to carry out multi-angle interactive preview.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence design technology, specifically to a method and system for personalized gift customization based on multimodal AIGC. Background Technology

[0002] Current personalized gift customization primarily relies on template selection and parameter adjustment. Existing technologies largely depend on single text descriptions or image references to understand user needs, making it difficult to comprehensively capture the user's multi-dimensional design intentions. The information acquisition process is fixed, failing to dynamically optimize sampling strategies based on input quality. Feature extraction methods are isolated, failing to effectively integrate information from different modalities. Design generation models are simple, typically based on generative adversarial networks or variational autoencoders, resulting in insufficient diversity and controllability of generated results. Image optimization methods are limited, often employing global filtering or simple super-resolution techniques, lacking targeted enhancement for key areas. Design output formats are fixed, making it difficult for users to flexibly adjust the generated results. Preview methods are traditional, with static images failing to accurately reflect the actual effect of the gift. Existing methods need to address key technical challenges such as multi-modal information fusion, accurate interpretation of design intentions, controllable optimization of the generation process, and improvement of the interactive experience.

[0003] Traditional gift customization systems suffer from significant shortcomings in terms of intelligence and user experience. Multimodal data alignment accuracy is low, and a semantic gap exists between text descriptions and reference images. The encoder structure is simple, mapping different modal features to different vector spaces, making deep interaction difficult. Intent parsing is linear, failing to capture users' implicit design preferences and emotional inclinations. Diffusion model training is unstable, and the denoising process is prone to mode collapse and loss of detail. Conditional control mechanisms are rigid, lacking fine-grained constraints on diversity adjustments. Content-aware algorithms are computationally complex, resulting in low real-time processing efficiency. Template conversion processes suffer severe information loss, limiting editability. Augmented reality rendering effects are distorted, and the integration of virtual and real scenes is unnatural. Existing technologies necessitate the establishment of a fully intelligent customization solution that spans the entire process from needs understanding to immersive preview. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for personalized gift customization based on multimodal AIGC, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a gift personalization customization method based on multimodal AIGC, the method comprising: The system acquires multimodal input information provided by the user, including text descriptions, reference images, style tags, and gift usage categories, and dynamically samples and optimizes the acquisition process. The multimodal input information is jointly encoded using a multimodal encoder to generate a multimodal feature embedding, and the user design intent vector is parsed out using an intent decoder. Based on the user design intent vector, a diffusion model is used to iteratively denoise and generate a draft image of the gift design, and a conditional control network is used to adjust the diversity of the generation process. The initial draft image of the gift design is reconstructed using a content-aware mechanism to output an optimized gift design image. The optimized gift design image is converted into an editable design template and integrated with an augmented reality engine for multi-angle interactive preview.

[0006] Preferably, the acquisition of user-provided multimodal input information further includes dynamic sampling optimization of the acquisition process, including: Monitor the data quality and integrity of the user input stream, and automatically adjust the sampling frequency and buffer size based on historical input patterns; Perform time-series alignment and missing value imputation on real-time input multimodal information; An importance-weighted algorithm is used to dynamically sample multimodal input information, prioritizing the retention of data segments with high information entropy.

[0007] Preferably, the multimodal input information is jointly encoded using a multimodal encoder to generate a multimodal feature embedding, and the user design intent vector is parsed using an intent decoder, including: Input the text description into the text encoder to extract semantic features, input the reference image into the visual encoder to extract visual features, and input the style tag and gift usage category into the category encoder to extract category features; By fusing the semantic features, visual features, and category features through a cross-modal attention mechanism, a unified multimodal feature embedding is generated; The multimodal feature embedding is sequentially decoded using an intent decoder to output a user design intent vector. The intent decoder is built based on the Transformer architecture.

[0008] Preferably, based on the user design intent vector, a diffusion model is used to iteratively denoise and generate a draft image of the gift design, and a conditional control network is used to adjust the diversity of the generation process, including: A noise prediction network for a diffusion model that uses user design intent vectors as conditional inputs; Multi-step iterative denoising is performed in the latent space to gradually restore the initial draft image of the gift design; Based on the initial draft image, the denoising step size and noise level are adjusted through a conditional control network to control the diversity and clarity of the generated image. Using peak signal-to-noise ratio as the convergence criterion, the iteration process is terminated and the initial draft image of the gift design is output.

[0009] Preferably, the initial draft image of the gift design is subjected to local detail reconstruction using a content-aware mechanism to output an optimized gift design image, including: Extract content and style features from the initial draft images of the gift design, and then integrate the target style using a style transfer network; The content-aware mechanism is used to identify key regions in an image, and feature enhancement and resolution enhancement are performed on these key regions. By using a detail reconstruction network to repair blurry or broken parts of an image, an optimized gift design image with coherent texture is generated.

[0010] Preferably, the optimized gift design image is converted into an editable design template, and an augmented reality engine is integrated for multi-angle interactive preview, including: We will optimize the vectorization of gift design images to generate layered, editable design templates. Import the layered, editable design template into the augmented reality engine and bind physical material properties and lighting models; The gesture recognition interface enables users to rotate, scale, and change materials on design templates.

[0011] Preferably, the semantic features, visual features, and category features are fused through a cross-modal attention mechanism to generate a unified multimodal feature embedding, including: Calculate the attention weights between semantic and visual features to highlight cross-modal associations; Injecting categorical features as bias terms into the attention layer enhances the category orientation of feature fusion; Gated recurrent units are used to temporally smooth the fused features, generating stable multimodal feature embeddings.

[0012] Preferably, a content-aware mechanism is used to identify key regions in an image, and feature enhancement and resolution improvement are performed on these key regions, including: A saliency detection algorithm is used to locate the visual focal region in the image, and a super-resolution network is applied to the visual focal region to improve the detail resolution.

[0013] Preferably, the method further includes: collecting user interaction behavior data through a feedback system, dynamically updating the parameters of the intent decoder and the diffusion model, and realizing real-time optimization of the generation strategy.

[0014] Preferably, the present invention also includes a gift personalization system based on multimodal AIGC, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described gift personalization method based on multimodal AIGC.

[0015] Compared with the prior art, the beneficial effects of the present invention are: A multimodal encoder is used to jointly encode multimodal input information, including text descriptions, reference images, style tags, and gift usage categories, generating multimodal feature embeddings. An intent decoder then parses these embeddings to extract the user's design intent vector. The multimodal encoder employs a cross-attention mechanism to achieve interaction and alignment between features from different modalities. Semantic features of the text description are extracted using a pre-trained language model, while visual features of the reference image are extracted using a convolutional neural network. Style tags and usage categories are converted into vector representations through embedding layers. The joint encoding process considers intermodal complementarity, eliminating information redundancy and conflict. The multimodal feature embeddings are represented in a shared latent space, preserving the core information of each modality. The intent decoder uses a recurrent neural network or transformer structure to decode the user's design intent from the feature embeddings. The design intent vector comprehensively reflects the user's aesthetic preferences, functional needs, and emotional appeals. The vector dimensions are optimized to balance expressive power and computational efficiency. The parsing process uses an attention mechanism to focus on key features, improving the accuracy of intent extraction.

[0016] Based on the user's design intent vector, a diffusion model is used iteratively to denoise and generate initial draft images of gift designs. A conditional control network adjusts the diversity of the generation process. The diffusion model progressively adds noise in the forward pass and progressively denoises in the reverse pass. The conditional control network uses the design intent vector as a conditional input to guide the image generation direction. The iterative denoising process uses stochastic differential equations to ensure the stability of the generation process. Diversity adjustment is achieved by adjusting the noise injection intensity and sampling step size. The conditional control network uses an adaptive normalization layer to inject conditional information into each generation step. The generation process employs either classifier-guided or classifier-free strategies to balance generation quality and diversity. The initial draft images contain basic composition, color scheme, and stylistic features. The generated results undergo quality evaluation, filtering out low-quality samples. A content-aware mechanism is applied to the initial draft images of the gift designs for local detail reconstruction, outputting optimized gift design images. The content-aware mechanism locates key regions using saliency detection and edge recognition algorithms. Local detail reconstruction uses an adversarial generative network or a diffusion model for region optimization. Detail enhancement considers texture coherence and visual rationality, avoiding artificial artifacts. The optimization process employs a progressive strategy, gradually improving image resolution and quality. Content-aware weights are dynamically allocated based on regional importance, with key regions receiving more optimization resources. The optimized image maintains stylistic consistency while enhancing visual appeal. Image quality assessment utilizes a multi-objective optimization approach combining perceptual and adversarial loss.

[0017] The process transforms optimized gift design images into editable design templates and integrates an augmented reality engine for multi-angle interactive previews. Template conversion is achieved through image segmentation and vectorization, preserving layer and path information. Editable design templates support color adjustments, element replacement, and layout modifications. The template structure is layered for easy user personalization. The augmented reality engine captures the real-world environment through a camera, overlaying the virtual design onto the real scene. Multi-angle interactive previews support rotation, scaling, and lighting adjustments. Real-time rendering during preview ensures smooth interaction. The virtual-real fusion utilizes lighting estimation and shadow generation technologies to enhance scene realism. Users can operate via gestures and voice commands, improving the naturalness of the interaction. Through the synergy of multimodal fusion, intent parsing, conditional generation, and augmented preview, the gift customization process is made intelligent. Multimodal coding comprehensively understands user needs, intent parsing accurately captures design concepts, conditional generation ensures product quality, and augmented preview enhances the user experience. This integrated approach significantly improves the efficiency and quality of personalized gift customization. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the gift personalization customization method based on multimodal AIGC described in this invention. Figure 2 A flowchart for dynamic sampling optimization for acquiring multimodal input information; Figure 3 A flowchart for joint encoding of multimodal input information and parsing of user intent vectors; Figure 4 A quality analysis diagram of the diffusion model iteration process; Figure 5 A comparative analysis chart of the migration effects of designs with multiple styles. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1This invention provides a method and system for personalized gift customization based on multimodal AIGC. The method includes: acquiring multimodal input information provided by the user, including text description, reference image, style tags, and gift usage category, and dynamically sampling and optimizing the acquisition process; jointly encoding the multimodal input information using a multimodal encoder to generate multimodal feature embeddings, and parsing the user's design intent vector using an intent decoder; generating a draft gift design image based on the user's design intent vector using a diffusion model for iterative denoising, and adjusting the diversity of the generation process using a conditional control network; applying a content-aware mechanism to reconstruct local details on the draft gift design image, and outputting an optimized gift design image; converting the optimized gift design image into an editable design template, and integrating an augmented reality engine for multi-angle interactive preview.

[0021] Example 1: See Figure 2 In specific implementation, monitoring the data quality and integrity of the user input stream includes real-time calculation of the signal-to-noise ratio and missing rate of multimodal input information, automatic adjustment of sampling frequency and buffer size based on historical input patterns (obtained through sliding window analysis of user-provided text descriptions, reference images, style tags, and gift usage categories), and dynamic configuration of system parameters using a time-series prediction-based algorithm to automatically adjust sampling frequency and buffer size. In some embodiments, time-series alignment and missing value imputation are performed on the real-time input multimodal information. Time-series alignment unifies the data streams of text descriptions, reference images, style tags, and gift usage categories onto a common time axis through a timestamp synchronization mechanism. Missing value imputation uses a linear regression model to fill in blank data points in the multimodal input information. It can be understood that an importance-weighted algorithm is used to dynamically sample the multimodal input information. The importance-weighted algorithm calculates the importance weight of each data segment based on information entropy, prioritizing the retention of data segments with high information entropy. Data segments with high information entropy are defined as multimodal input information parts with higher uncertainty and information content. Optionally, the importance-weighted algorithm uses the following formula to calculate the weights:

[0022] in: This represents the importance weight of the i-th data segment. This represents the information entropy of the i-th data segment. This represents the total number of data segments in the multimodal input information. In specific implementations, the dynamic sampling process filters multimodal input information in real time based on importance weights, ensuring that high-information-entropy data segments are preferentially retained for subsequent processing. In some embodiments, monitoring the data quality and integrity of the user input stream also involves using a quality control module to verify the format consistency and content validity of the multimodal input information, automatically adjusting the sampling frequency and buffer size, and optimizing system resource allocation through a feedback loop. Optionally, temporal alignment and missing value imputation are integrated with multimodal data fusion techniques to ensure consistency of text descriptions, reference images, style tags, and gift usage categories over time.

[0023] Example 2: See Figure 3 In specific implementation, the text description is input into a text encoder to extract semantic features. The text encoder uses a Transformer-based pre-trained language model to process the user-provided text description. A reference image is input into a visual encoder to extract visual features. The visual encoder uses a convolutional neural network architecture to process the user-provided reference image. Style tags and gift usage categories are input into a category encoder to extract category features. The category encoder maps discrete category information into dense vector representations through an embedding layer. A cross-modal attention mechanism is used to fuse semantic, visual, and category features to generate a unified multimodal feature embedding. Attention weights between semantic and visual features are calculated to highlight cross-modal associations. Category features are injected as a bias term into the attention layer to enhance the category-oriented nature of feature fusion. In some embodiments, the cross-modal attention mechanism calculates the association strength between semantic and visual features through query-key attention. The calculation process of attention weights is expressed by the following formula:

[0024] in: This represents the attention weight between the i-th semantic feature and the j-th visual feature. This represents the similarity score between semantic features and visual features. This represents the length of the visual feature sequence. A gated recurrent unit (GRU) is used to temporally smooth the fused features, generating a stable multimodal feature embedding. The GRU controls the flow of feature information through update and reset gates. It can be understood that an intent decoder is used to sequentially decode the multimodal feature embedding, outputting a user design intent vector. The intent decoder is built on a Transformer architecture, progressively generating the user design intent vector through a multi-layer self-attention mechanism and a feedforward neural network. Optionally, categorical features are injected as bias terms into the attention layer. Specifically, a linear transformation projects the categorical features onto the attention distribution space, and the projected categorical feature vector is added to the original attention distribution to achieve feature enhancement. In some embodiments, when calculating the attention weights between semantic and visual features, the similarity between the semantic feature vector and the visual feature vector is calculated through a dot product operation to obtain the attention distribution. It can be understood that when the GRU temporally smooths the fused features, it uses a gating mechanism to control the fusion ratio of historical and current feature information, ensuring the stability of the multimodal feature embedding. Optionally, the intent decoder is built on the Transformer architecture and includes an encoder-decoder structure. It processes multimodal feature embedding sequences through a multi-head self-attention mechanism and gradually decodes to generate user design intent vectors.

[0025] Example 3: In a specific implementation, the user design intent vector is used as a conditional input to the noise prediction network of the diffusion model. The noise prediction network uses a U-Net architecture to process noise data in the latent space, performing multi-step iterative denoising in the latent space to gradually reconstruct the initial draft image of the gift design. The iterative denoising process achieves image reconstruction by predicting the noise at the current time step and gradually subtracting the noise. Based on the initial draft image, a conditional control network adjusts the denoising step size and noise level to control the diversity and sharpness of the generated image. The conditional control network dynamically adjusts the sampling process of the diffusion model through a learnable parameter matrix. The peak signal-to-noise ratio (PSNR) is used as the convergence criterion to terminate the iterative process and output the initial draft image of the gift design. The PSNR is calculated based on the mean square error between the original image and the generated image. In some embodiments, the process of the conditional control network adjusting the denoising step size adopts an adaptive strategy, dynamically adjusting the number of subsequent denoising steps according to the image quality index. The noise level is adjusted by modifying the noise scheduling parameters in the diffusion model. It is understandable that a saliency detection algorithm is used to locate the visual focal region in an image. This algorithm identifies the distribution of high-frequency components in the image based on frequency domain analysis. A super-resolution network is then applied to the visual focal region to enhance detail resolution. The super-resolution network uses a residual dense block structure to enhance local texture details. Optionally, the conditional control network adjusts the diversity of the generation process according to the following formula:

[0026] in: This represents the noise level coefficient after adjustment at step t. Indicates diversity control factor. This represents the baseline noise level at step t. This represents the cumulative product coefficient at step t. In specific implementations, the multi-step iterative denoising process uses a stochastic differential equation solver to numerically solve the backdiffusion process, gradually restoring image feature information. In some embodiments, the saliency detection algorithm determines the visual focus position by calculating the contrast intensity of the image region, and the super-resolution network achieves resolution improvement through sub-pixel convolutional layers. It can be understood that when the peak signal-to-noise ratio (PSNR) is used as a convergence criterion, a threshold range is set, and the iteration process terminates when the change in PSNR over three consecutive iterations is less than the set threshold. Optionally, the noise prediction network's operation in the latent space projects high-dimensional image data into a low-dimensional latent space, reducing computational complexity while maintaining the integrity of image features.

[0027] See Figure 4 This chart illustrates the changes in image quality during the generation of a gift design draft based on multi-step iterative denoising. The chart uses a dual Y-axis design: the left axis displays the peak signal-to-noise ratio (PSNR), measuring the similarity between the generated and target images; the right axis shows the trend of the noise level coefficient. The solid blue line represents the gradual improvement in image quality as the number of iterations increases, showing a significant improvement in the initial stage followed by stabilization. The dashed red line illustrates the noise level decay process, demonstrating the core mechanism of the diffusion model in gradually removing noise and restoring image features. The dashed green line indicates the preset convergence threshold; when the image quality reaches this threshold, the system automatically terminates the iteration process, ensuring a balance between generation efficiency and quality. The entire chart clearly presents the technical characteristics of the diffusion model in multi-step iterative denoising in the latent space, and the optimization effect of the conditional control network on the generation process by adjusting the noise level and denoising step size.

[0028] Example 4: In specific implementation, the content and style features of the initial draft image of the gift design are extracted using a pre-trained convolutional neural network. Content features are extracted from the deep convolutional layers of the network, while style features are calculated from the statistical features of multiple convolutional layers. The target style is fused through a style transfer network, which uses an adaptive instance normalization method to align the statistical distributions of content and style features. A content-aware mechanism is used to identify key regions in the image. This mechanism calculates the spatial weight distribution of feature maps using a convolutional attention module, enhancing features and improving resolution in key regions. Feature enhancement employs a channel attention mechanism to strengthen the response of important feature channels, while resolution improvement uses spatial pyramid pooling to fuse multi-scale feature information. A detail reconstruction network repairs blurred or broken parts in the image, generating an optimized gift design image with coherent texture. The detail reconstruction network uses an encoder-decoder structure combined with skip connections to maintain high-frequency detail information. The optimized gift design image is vectorized to generate a layered, editable design template. Vectorization uses an image segmentation algorithm based on edge detection and region growing. The layered, editable design template contains multiple independent editing layers. Layered, editable design templates are imported into the augmented reality engine and bound to physically based material properties and lighting models. Physical material properties include diffuse reflection coefficient, specular intensity, and roughness parameters, while the lighting model uses physically based rendering technology to simulate realistic lighting effects. A gesture recognition interface allows users to rotate, scale, and replace materials on the design template. This interface uses a depth camera to capture the movement trajectory of key points on the user's hand.

[0029] In some embodiments, the calculation process for identifying key regions by the content-aware mechanism uses the following formula:

[0030] in: This represents the attention weight at position (x, y). This represents the learnable temperature coefficient. This represents the eigenvector at position (x, y). and These represent the height and width of the feature map, respectively. It can be understood that when the style transfer network fuses the target style, it optimizes the feature representation by calculating the difference between the Gram matrix of content features and style features. Adaptive instance normalization aligns the mean and variance of content features with the statistical properties of style features. Optionally, referring to Table 1, the detail reconstruction network uses a gated convolution mechanism to control the feature propagation path when repairing image defects, and skip connections fuse the shallow features of the encoder with the deep features of the decoder to maintain detail integrity.

[0031] Table 1: Data Structure Table of Layered Editable Design Template

[0032] In some embodiments, when binding physical material properties, corresponding material parameters are assigned according to the semantic type of the design element, and the lighting model generates realistic shadow effects through ambient occlusion and global illumination calculations. It can be understood that the gesture recognition interface maps the user's hand movement trajectory into a transformation matrix in three-dimensional space. Rotation operations calculate the hand rotation angle through the gesture recognition interface, scaling operations adjust the model size according to changes in the distance between the hands, and material replacement operations trigger material library switching through preset gestures. Optionally, during vectorization, Bézier curves are used to fit the image contours, layered editable design templates store layer relationships and editing attributes in SVG format, and the augmented reality engine overlays the design template onto the real environment using spatial anchor point positioning technology.

[0033] See Figure 5 The system compared key performance indicators (KPIs) of five different design styles during the transfer process, including modern minimalism, traditional classicism, natural freshness, futuristic technology, and artistic abstraction. The charts, presented as grouped bar charts, displayed data across three dimensions: enhancement ratio, final quality score, and attention intensity. The enhancement ratio reflects the degree of matching between content features and style features, demonstrating the effectiveness of the style transfer network in aligning statistical distributions through adaptive instance normalization. The final quality score comprehensively evaluates the visual quality and detail integrity of the generated image, showcasing the role of the detail reconstruction network in repairing image defects and enhancing texture coherence. The attention intensity indicator demonstrates the content-aware mechanism's ability to identify key image regions; the spatial weight distribution calculated by the convolutional attention module effectively improves the feature representation of important regions. Different colored bars visually compare the performance differences of various styles across different KPIs, providing data support for design optimization.

[0034] Example 5: In a specific implementation, user interaction behavior data is collected through a feedback system. The feedback system records the user's rotation angle, scaling ratio, and material replacement frequency data for the gift design template during the augmented reality preview stage. This user interaction behavior data is stored in a behavior log database in time-series format. The parameters of the intent decoder and diffusion model are dynamically updated using an online learning algorithm, which optimizes model weights based on stochastic gradient descent. In some embodiments, the feedback system simultaneously records the final adoption status of the design template while collecting user interaction behavior data. Explicit user ratings and implicit operation durations are used as important indicators for optimization. The parameters of the intent decoder and diffusion model are dynamically updated using a mini-batch training strategy, sampling training samples from the behavior log database in chronological order. It can be understood that real-time optimization of the generation strategy is achieved by establishing a reward function to evaluate the generation effect. The reward function comprehensively calculates the user satisfaction index and the design consistency score. The parameter update process follows the following formula:

[0035] in: This indicates the amount of time the model parameters have been updated. Indicates the dynamic learning rate. Represents the loss function For parameters gradient, This indicates a reward signal from the feedback system. This represents the model's predicted output. In specific implementations, the attention weight matrix in the Transformer network is adjusted when dynamically updating the intent decoder parameters, and the convolution kernel parameters in the noise prediction network are corrected when dynamically updating the diffusion model parameters. Optionally, the feedback system collects user interaction behavior data, including view dwell time, edit operation sequences, and design element modification history. This behavior data is standardized and used as training samples for online learning. In some embodiments, the policy gradient method is used to adjust the output distribution of the conditional control network when optimizing the generation strategy in real time, and the parameter updates of the intent decoder are concentrated on the projection matrix of the cross-modal attention layer. The specific implementation of the policy gradient method includes constructing a reward function based on the user interaction behavior data collected by the feedback system. The reward function integrates the user satisfaction index and the design consistency score. The user satisfaction index comes from the final adoption status of the design template and the implicit operation duration, and the design consistency score is calculated by comparing the semantic alignment between the generated image and the user's multimodal input information. It can be understood that the dynamic update process uses a sliding window mechanism to process the data flow, ensuring the timeliness of the training data and the stability of the model parameters. The feedback system processes multi-user interaction data in parallel through a distributed computing architecture. Optionally, the online learning algorithm sets a loss threshold to trigger parameter updates. When the average loss of consecutive batches exceeds the preset threshold, the parameter adjustment process of the intent decoder and diffusion model is initiated. The calculation of the reward signal integrates user operation behavior and the adoption decision of design results.

[0036] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for personalized gift customization based on multimodal AIGC, characterized in that, The method includes the following steps: The system acquires multimodal input information provided by the user, including text descriptions, reference images, style tags, and gift usage categories, and dynamically samples and optimizes the acquisition process. The multimodal input information is jointly encoded using a multimodal encoder to generate a multimodal feature embedding, and the user design intent vector is parsed out using an intent decoder. Based on the user design intent vector, a diffusion model is used to iteratively denoise and generate a draft image of the gift design, and a conditional control network is used to adjust the diversity of the generation process. The initial draft image of the gift design is reconstructed using a content-aware mechanism to output an optimized gift design image. The optimized gift design image is converted into an editable design template and integrated with an augmented reality engine for multi-angle interactive preview.

2. The gift personalization customization method based on multimodal AIGC according to claim 1, characterized in that, The acquisition of user-provided multimodal input information also includes dynamic sampling optimization of the acquisition process, including: Monitor the data quality and integrity of the user input stream, and automatically adjust the sampling frequency and buffer size based on historical input patterns; Perform time-series alignment and missing value imputation on real-time input multimodal information; An importance-weighted algorithm is used to dynamically sample multimodal input information, prioritizing the retention of data segments with high information entropy.

3. The gift personalization method based on multimodal AIGC according to claim 1, characterized in that, The multimodal input information is jointly encoded using a multimodal encoder to generate a multimodal feature embedding, and the user's design intent vector is parsed using an intent decoder, including: Input the text description into the text encoder to extract semantic features, input the reference image into the visual encoder to extract visual features, and input the style tag and gift usage category into the category encoder to extract category features; By fusing the semantic features, visual features, and category features through a cross-modal attention mechanism, a unified multimodal feature embedding is generated; The multimodal feature embedding is sequentially decoded using an intent decoder to output a user design intent vector. The intent decoder is built based on the Transformer architecture.

4. The gift personalization customization method based on multimodal AIGC according to claim 1, characterized in that, Based on the user design intent vector, a diffusion model is used for iterative denoising to generate a draft image of the gift design. A conditional control network is then used to adjust the diversity of the generation process, including: A noise prediction network for a diffusion model that uses user design intent vectors as conditional inputs; Multi-step iterative denoising is performed in the latent space to gradually restore the initial draft image of the gift design; Based on the initial draft image, the denoising step size and noise level are adjusted through a conditional control network to control the diversity and clarity of the generated image. Using peak signal-to-noise ratio as the convergence criterion, the iteration process is terminated and the initial draft image of the gift design is output.

5. The gift personalization customization method based on multimodal AIGC according to claim 1, characterized in that, The initial draft image of the gift design is reconstructed using a content-aware mechanism to extract local details, resulting in an optimized gift design image. This includes: Extract content and style features from the initial draft images of the gift design, and then integrate the target style using a style transfer network; The content-aware mechanism is used to identify key regions in an image, and feature enhancement and resolution improvement are performed on these key regions. By using a detail reconstruction network to repair blurry or broken parts of an image, an optimized gift design image with coherent texture is generated.

6. The gift personalization method based on multimodal AIGC according to claim 1, characterized in that, The optimized gift design image is converted into an editable design template, and an augmented reality engine is integrated for multi-angle interactive preview, including: We will optimize the vectorization of gift design images to generate layered, editable design templates. Import the layered, editable design template into the augmented reality engine and bind physical material properties and lighting models; The gesture recognition interface enables users to rotate, scale, and change materials on design templates.

7. The gift personalization customization method based on multimodal AIGC according to claim 3, characterized in that, By fusing the semantic, visual, and category features through a cross-modal attention mechanism, a unified multimodal feature embedding is generated, including: Calculate the attention weights between semantic and visual features to highlight cross-modal associations; Injecting categorical features as bias terms into the attention layer enhances the category orientation of feature fusion; Gated recurrent units are used to temporally smooth the fused features, generating stable multimodal feature embeddings.

8. The gift personalization method based on multimodal AIGC according to claim 5, characterized in that, Content-aware mechanisms are used to identify key regions in an image, and feature enhancement and resolution enhancement are performed on these key regions, including: A saliency detection algorithm is used to locate the visual focal region in the image, and a super-resolution network is applied to the visual focal region to improve the detail resolution.

9. The gift personalization method based on multimodal AIGC according to claim 1, characterized in that, The method further includes: collecting user interaction behavior data through a feedback system, dynamically updating the parameters of the intent decoder and diffusion model, and realizing real-time optimization of the generation strategy.

10. A gift personalization system based on multimodal AIGC, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the gift personalization customization method based on multimodal AIGC as described in any one of claims 1 to 9.

Images