Product design method based on generative artificial intelligence

Abstract

The invention discloses a product design method based on generative artificial intelligence, and the method comprises the steps: training a pattern image generation model: carrying out the adapter training of pattern image features through employing an image generation model FLUX 1.0 as a basic model based on a preprocessed pattern image data set, and employing a low-rank adaptive method LoRA; the method specifically comprises the steps that under the condition that the weight of a basic model is frozen, low-rank matrix decomposition is introduced to restrain and represent the model parameter update amount, and a low-rank matrix is initialized to reduce initial disturbance on a pre-training model, so that visual features and cultural semantics of a pattern image are embedded into the basic model; and building a product design workflow: building the product design workflow based on a generative artificial intelligence creation framework ComfyUI and a trained image generation model, and realizing real-time generation of a product. The method has the advantages that cultural semantics and composition features can be understood, and the generated image is high in accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent design, and more specifically to a product design method based on generative artificial intelligence. Background Technology

[0002] In recent years, groundbreaking advancements in artificial intelligence have significantly propelled the development of conventional workflow technologies. These technologies can be used to constrain the generation process of AI models, preventing illusions. Within a workflow framework, the collaborative operation of multiple models can be efficiently achieved, thereby greatly improving the efficiency and controllability of complex tasks. Therefore, workflow construction based on node-based, visual programming frameworks has attracted considerable attention. Among them, ComfyUI, as an open-source, highly flexible, and powerful workflow engine, provides a robust practical platform for AI generation.

[0003] In existing technologies, Xiangyuan Xue et al. proposed ComfyBench to evaluate the ability of LLM agents to design collaborative AI systems in ComfyUI. Based on this, they developed the ComfyAgent framework, which can represent ComfyUI workflows in code, enabling LLMs to plan and understand workflows at the code level. ZhenranXu et al. proposed ComfyUI-R1, a large-scale inference model designed specifically for automated workflow generation, aiming to give the model inference and planning capabilities. Oucheng Huang et al. proposed ComfyGPT, a self-optimizing multi-agent system that can automatically generate ComfyUI workflows based on task descriptions and connect nodes through single links, revolutionizing the process generation method and significantly improving LLMs' ability to handle complex interdependencies in ComfyUI topology workflow generation. However, when the generated objects focus on culturally rich carriers such as batik patterns and cultural and creative products, the limitations of existing workflow frameworks in lacking understanding of the cultural semantics and compositional features of batik art become apparent, making it difficult to ensure the cultural accuracy of the generated results. Furthermore, most existing workflows focus on the single goal of pattern generation, failing to achieve a complete design process from pattern generation to the realization of cultural and creative products, thus limiting the application of generative artificial intelligence in the digital innovation of intangible cultural heritage. Summary of the Invention

[0004] To overcome the above-mentioned shortcomings, this invention proposes a product design method based on generative artificial intelligence that can understand cultural semantics and compositional features and generate images with high accuracy.

[0005] The present invention provides a product design method based on generative artificial intelligence, wherein the method includes the following steps:

[0006] Step 1: Construction and preprocessing of pattern image dataset: Collect pattern image data and perform preprocessing;

[0007] Step 2: Pattern Image Generation Model Training: Based on the preprocessed pattern image dataset, using the image generation model FLUX 1.0 as the base model, the low-rank adaptive method LoRA is employed for adapter training of pattern image features. The specific process includes: freezing the weights of the base model, introducing low-rank matrix factorization to constrain and represent the parameter update amount of the model, and initializing the low-rank matrix to reduce initial perturbations to the pre-trained model, thus embedding the visual features and cultural semantics of the pattern images into the base model. After training, a dedicated pattern image LoRA adapter node is obtained, enabling controllable generation of this type of pattern image.

[0008] Step 3: Building the Product Design Workflow: Based on the generative AI creation framework ComfyUI and the trained image generation model, the product design workflow is built. Specific steps include:

[0009] Step 3.1: Integrate a high-efficiency inference engine to optimize generation efficiency: The product design flow integrates the Nunchaku high-efficiency inference engine for diffusion models to optimize nodes;

[0010] Step 3.2: Pattern feature mapping based on multimodal semantic alignment: Perform contrastive language-image preprocessing training CLIP at the front end of the workflow to construct a cross-modal semantic alignment channel between text and pattern images;

[0011] The contrastive language-image preprocessing training CLIP is used to convert user-input images and text into feature vectors that the model can understand, specifically including:

[0012] First, in the contrastive language-image preprocessing training CLIP, the text encoder transforms the abstract description of the user-input text into a high-dimensional semantic vector. , In the field of pattern image processing, textual semantic feature vectors aligned with the visual feature space are constructed by analyzing technical terms and cultural symbolic metaphors. An image encoder, VisionTransformer, is then used to encode the pattern image into a visual feature vector of the same dimension. , Image visual feature vectors representing pattern color contrast and pattern structural complexity;

[0013] Then, the cosine similarity between the text semantic feature vector and the image visual feature vector is calculated. ;

[0014] Step 3.3: Apply the pattern image generation model trained in Step 2, and through the dedicated pattern image LoRA adapter node, achieve precise control and generation of pattern features;

[0015] Step 3.4: Real-time Product Generation: The pattern image output from Step 3.3 is used as input. At the same time, the semantic description of product type, material properties and usage scenario is parsed through CLIP text encoder, and the product white model LoRA is loaded to maintain the geometric integrity of the product. A cross-attention mechanism is used to achieve spatial alignment between pattern features and product structure. Combined with Nunchaku's 4-bit quantization acceleration and progressive generation strategy, high-fidelity fusion of pattern and product appearance is achieved to generate product design scheme image.

[0016] The above-mentioned product design method based on generative artificial intelligence, wherein: in step 1, the collected pattern image data is a batik pattern image with a resolution of 1024×1024.

[0017] The above-mentioned product design method based on generative artificial intelligence includes, in step 1, the preprocessing including brightness adjustment, color adjustment, and normalization processing.

[0018] The above-mentioned product design method based on generative artificial intelligence includes the following step: In step 2, the low-rank adaptive method involves LoRA training using the Kohya_ss toolbox; during training, the parameter update matrix of the pre-trained language model is obtained by decomposing the weight matrix. Designed as a low-rank matrix and The product of the two weights preserves the basic weight matrix. Frozen state, only through low-rank matrix and To implement parameter updates, it is represented as follows:

[0019]

[0020]

[0021] Wherein, the original weight matrix , Represents the original weight matrix the number of rows, Represents the original weight matrix The number of columns.

[0022] The above-mentioned product design method based on generative artificial intelligence, wherein: in step 2, the low-rank matrix By reconstructing the low-rank subspace features to the dimension of the basic weight matrix, we can deeply analyze the knowledge of pattern craftsmanship, understand the metaphorical system and cultural connotations of the patterns, and thus... is the rank parameter, representing the dimension of the low-rank subspace.

[0023] The above-mentioned product design method based on generative artificial intelligence, wherein: the rank parameter ,satisfy Less than or equal to and The minimum value, i.e. The value can be 1, 2, 4 or 8.

[0024] The above-mentioned product design method based on generative artificial intelligence, wherein: in step 2, the low-rank matrix The pattern image features input from the dataset are mapped to a low-rank subspace, thereby capturing the shallow set topological features of the patterns.

[0025] The above-mentioned product design method based on generative artificial intelligence, wherein: in step 2, the low-rank matrix and Initialization: Low-rank matrix Using a mean of 0 and a standard deviation of , Gaussian distribution initialization This allows for the introduction of controllable randomness to explore potential combination patterns of pattern features; low-rank matrix Then initialize it as a zero matrix. This ensures the initial training phase This preserves the original generation and understanding capabilities of the model, avoiding damage to the general representation capabilities of the trained model.

[0026] The above-mentioned product design method based on generative artificial intelligence, wherein: in step 2, the low-rank matrix and The projection matrix of query, key, and value in the cross-attention module U-Net applied to the image generation model FLUX 1.0. , , In the formula, it is:

[0027] ;

[0028] ;

[0029] ;

[0030] in, It is a query matrix. It is a key matrix. It is a value matrix. These are the original weights of the Query projection matrix in the pre-trained model, which map each token in the input sequence to the query space and are used to calculate the attention score. These are the original weights of the Key projection matrix, which map each token of the input sequence to the key space for matching in queries; These are the original weights of the Value projection matrix in the pre-trained model, which map each token of the input sequence to the value space and are used to generate the attention-weighted context representation. and It is a low-rank matrix and The parameter matrix corresponding to the Query projection matrix; and It is a low-rank matrix and The parameter matrix corresponding to the Key projection matrix; and It is a low-rank matrix and The parameter matrix corresponding to the Value projection matrix.

[0031] The above-mentioned product design method based on generative artificial intelligence, wherein: in step 3.2, the cosine similarity between the text semantic feature vector and the image visual feature vector is calculated. , is represented as:

[0032]

[0033] in, , The closer to 1, the more likely the two vectors are in the same direction; the closer to -1, the more likely they are in opposite directions. When , it indicates orthogonality, meaning there is no correlation.

[0034] Compared with existing technologies, this invention has significant advantages. As can be seen from the above scheme, this invention provides a product design method based on generative artificial intelligence. Based on the ComfyUI framework, it provides an intelligent generation workflow for batik patterns and cultural and creative products, modularizing the generation process of patterns and cultural and creative products. It combines low-rank LoRA and CLIP nodes to constrain the generated content. The advantage of ComfyUI lies in decomposing the underlying operations of Stable Diffusion into separate, visual functional nodes. This architectural design solves the shortcomings of the fixed operation mode of Stable Diffusion, allowing operators to break down complex generation processes into sub-processes, thereby improving the accuracy and stability of pattern generation, optimizing generation efficiency, and reducing hardware burden. This research, through the integration of highly customizable node design, not only overcomes the deficiencies of Stable Diffusion in terms of operational flexibility and generation efficiency, but also achieves intelligent generation of batik patterns and cultural and creative product structures, styles, colors, and technological characteristics.

[0035] The following specific embodiments further illustrate the beneficial effects of the present invention. Attached Figure Description

[0036] Figure 1 This is a flowchart of the present invention;

[0037] Figure 2 This is the image generation result of the batik pattern cultural and creative product in the embodiment of the present invention. Detailed Implementation

[0038] The following detailed description, in conjunction with the accompanying drawings and preferred embodiments, outlines the specific implementation methods, features, and effects of a product design method based on generative artificial intelligence proposed in accordance with the present invention.

[0039] like Figure 1 As shown, the present invention provides a product design method based on generative artificial intelligence, wherein the method includes the following steps:

[0040] Step 1: Construction and preprocessing of pattern image dataset: Collect pattern image data and perform preprocessing;

[0041] The collected pattern image data are batik pattern images with a resolution of 1024×1024.

[0042] The preprocessing includes brightness adjustment, color adjustment, and normalization.

[0043] Step 2: Pattern Image Generation Model Training: Based on the preprocessed pattern image dataset, using the image generation model FLUX 1.0 as the base model, the low-rank adaptive method LoRA is employed for adapter training of pattern image features. The specific process includes: freezing the weights of the base model, introducing low-rank matrix factorization to constrain and represent the parameter update amount of the model, and initializing the low-rank matrix to reduce initial perturbations to the pre-trained model, thus embedding the visual features and cultural semantics of the pattern images into the base model. After training, a dedicated pattern image LoRA adapter node is obtained, enabling controllable generation of this type of pattern image.

[0044] The low-rank adaptive method involves LoRA training using the Kohya_ss toolbox; during training, the parameter update matrix of the pre-trained language model is obtained by decomposing the weight matrix. Designed as a low-rank matrix and The product of the two weights preserves the basic weight matrix. Frozen state, only through low-rank matrix and To implement parameter updates, it is represented as follows:

[0045]

[0046]

[0047] Wherein, the original weight matrix , Represents the original weight matrix the number of rows, Represents the original weight matrix The number of columns;

[0048] The low-rank matrix By reconstructing the low-rank subspace features to the dimension of the basic weight matrix, we can deeply analyze the knowledge of pattern craftsmanship, understand the metaphorical system and cultural connotations of the patterns, and thus... is the rank parameter, representing the dimension of the low-rank subspace.

[0049] The rank parameter ,satisfy Less than or equal to and The minimum value, i.e. The value can be 1, 2, 4 or 8.

[0050] The low-rank matrix The pattern image features input from the dataset are mapped to a low-rank subspace, thereby capturing the shallow set topological features of the patterns.

[0051] The low-rank matrix and Initialization: Low-rank matrix Using a mean of 0 and a standard deviation of , Gaussian distribution initialization This allows for the introduction of controllable randomness to explore potential combination patterns of pattern features; low-rank matrix Then initialize it as a zero matrix. This ensures the initial training phase This preserves the original generation and understanding capabilities of the model, avoiding damage to the general representation capabilities of the trained model.

[0052] The low-rank matrix and The projection matrix of query, key, and value in the cross-attention module U-Net applied to the image generation model FLUX 1.0. , , In the formula, it is:

[0053] ;

[0054] ;

[0055] ;

[0056] in, It is a query matrix. It is a key matrix. It is a value matrix. These are the original weights of the Query projection matrix in the pre-trained model, which map each token in the input sequence to the query space and are used to calculate the attention score. These are the original weights of the Key projection matrix, which map each token of the input sequence to the key space for matching in queries; These are the original weights of the Value projection matrix in the pre-trained model, which map each token of the input sequence to the value space and are used to generate the attention-weighted context representation. and It is a low-rank matrix and The parameter matrix corresponding to the Query projection matrix serves to specialize the semantic query mechanism of the model, focusing on visual features that conform to the batik craft and aesthetic system. and It is a low-rank matrix and The parameter matrix corresponding to the Key projection matrix is ​​responsible for reshaping the association and matching rules of the visual features inside the model, ensuring that the generated pattern conforms to the constraints and style of traditional craftsmanship in terms of structure. and It is a low-rank matrix and The parameter matrix corresponding to the Value projection matrix directly controls the visual presentation style of the final generated content, completing the stylization from abstract semantic description to specific batik images. Through the aforementioned low-rank correction, the response accuracy of the generative model to text description is enhanced, and its ability to represent fine-grained differences in input semantics is improved.

[0057] Step 3: Building the Product Design Workflow: Based on the generative AI creation framework ComfyUI and the trained image generation model, the product design workflow is built. Specific steps include:

[0058] Step 3.1: Integrate a high-efficiency inference engine to optimize generation efficiency: The product design flow integrates the Nunchaku high-efficiency inference engine for diffusion models to optimize nodes;

[0059] Step 3.2: Pattern feature mapping based on multimodal semantic alignment: Perform contrastive language-image preprocessing training CLIP at the front end of the workflow to construct a cross-modal semantic alignment channel between text and pattern images;

[0060] The contrastive language-image preprocessing training CLIP is used to convert user-input images and text into feature vectors that the model can understand, specifically including:

[0061] First, in the contrastive language-image preprocessing training CLIP, the text encoder transforms the abstract description of the user-input text into a high-dimensional semantic vector. , In the field of pattern image processing, textual semantic feature vectors aligned with the visual feature space are constructed by analyzing technical terms and cultural symbolic metaphors. An image encoder, VisionTransformer, is then used to encode the pattern image into a visual feature vector of the same dimension. , Image visual feature vectors representing pattern color contrast and pattern structural complexity;

[0062] Then, the cosine similarity between the text semantic feature vector and the image visual feature vector is calculated. , is represented as:

[0063]

[0064] in, , The closer to 1, the more likely the two vectors are in the same direction; the closer to -1, the more likely they are in opposite directions. When , it indicates orthogonality, meaning there is no correlation;

[0065] Step 3.3: Apply the pattern image generation model trained in Step 2, and through the dedicated pattern image LoRA adapter node, achieve precise control and generation of pattern features;

[0066] Step 3.4: Real-time Product Generation: Using the pattern image output from Step 3.3 as input, the semantic descriptions of product type, material properties, and usage scenarios are parsed using the CLIP text encoder, and the product white model LoRA is loaded to maintain the geometric integrity of the product. A cross-attention mechanism is used to achieve spatial alignment between pattern features and product structure. Combined with Nunchaku's 4-bit quantization acceleration and progressive generation strategy, high-fidelity fusion of the pattern and product appearance is achieved at a resolution of 1024×1024 to generate a product solution image.

[0067] Effect Analysis

[0068] To verify the effectiveness of this invention, a comparative experiment was conducted on a single NVIDIA RTX 4070Ti SUPER computer. The workflow for generating batik pattern images using LoRA trained with SDXL and Qwen-Image as the base model was compared with the workflow of this invention, which uses LoRA trained with Flux as the base model and has identical parameters. The experiment shows that this invention offers a certain improvement over existing methods.

[0069] To evaluate the quality of the pattern images generated by the workflow proposed in this invention, this invention employs a variety of quantitative indicators, including PSNR (Peak Signal-to-Noise Ratio) and SSIM (structural similarity index), to quantitatively analyze the quality and effect of the generated pattern images.

[0070] Peak Signal-to-Noise Ratio (PSNR) is a common objective quality evaluation metric in digital image processing. It assesses the degree of image distortion by comparing the original image with the distorted image. A higher PSNR value indicates less image distortion and better image quality. The calculation formula is:

[0071]

[0072] In the formula, MAX I The dynamic range of the image is represented by MSE, which measures the difference between the real batik pattern image and the generated pattern image. For an image of size [size missing], [the value missing]. The formula for calculating the MSE of grayscale images I and K is:

[0073]

[0074] The Structural Similarity Index (SSIM) is a metric used to measure the similarity between two images. It aims to assess image quality by comparing brightness, contrast, and structure. Unlike PSNR, which measures absolute error, SSIM is a perceptual model; its score is highly correlated with human visual perception of image similarity, accurately reflecting the similarity between two images. Its formula is:

[0075]

[0076] In the formula, Images of authentic batik patterns. To generate pattern images, and Representing images respectively and images The mean, and A constant added to avoid division by zero. The variance of the image. express and The covariance. The SSIM score ranges from 0 to 1, where 0 indicates that the two images are completely different, and 1 indicates that the two images are completely identical.

[0077] By performing quantitative analysis on the image of the generated pattern (such as...) Figure 2 The experimental results are shown in Table 1. The batik pattern image c(1) generated using Flux as the base model has the highest PSNR score, indicating that the image has the least distortion during the generation process, i.e., the highest image quality. The SSIM score is 0.79, indicating that the generated pattern is most similar to the real batik pattern. The image c(2) generated using SDXL as the base model has the lowest PSNR and SSIM scores, indicating the worst generation quality. Overall, the average PSNR score of the pattern generated by the Flux model is 28.4, and the average SSIM score is 0.75, which are the highest among the three models. The experimental results show that the batik pattern generation workflow based on the Flux model can generate new high-quality patterns similar to real batik patterns, and can maintain fine details and color coordination.

[0078] Table 1 Quantitative Evaluation Table for Batik Image Generation

[0079]

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A product design method based on generative artificial intelligence, characterized in that: The method includes the following steps: Step 1: Construction and preprocessing of pattern image dataset: Collect pattern image data and perform preprocessing; Step 2: Pattern Image Generation Model Training: Based on the preprocessed pattern image dataset, using the image generation model FLUX 1.0 as the base model, the low-rank adaptive method LoRA is used to train the adapter for pattern image features; The specific process includes: freezing the weights of the base model, introducing low-rank matrix factorization to constrain and represent the parameter update of the model, and initializing the low-rank matrix to reduce the initial perturbation to the pre-trained model, so that the visual features and cultural semantics of the pattern image are embedded into the base model; after training, a dedicated pattern image LoRA adapter node is obtained to realize the controllable generation of this type of pattern image. Step 3: Building the Product Design Workflow: Based on the generative AI creation framework ComfyUI and the trained image generation model, the product design workflow is built. Specific steps include: Step 3.1: Integrate a high-efficiency inference engine to optimize generation efficiency: The product design flow integrates the Nunchaku high-efficiency inference engine for diffusion models to optimize nodes; Step 3.2: Pattern feature mapping based on multimodal semantic alignment: Perform contrastive language-image preprocessing training CLIP at the front end of the workflow to construct a cross-modal semantic alignment channel between text and pattern images; The contrastive language-image preprocessing training CLIP is used to convert user-input images and text into feature vectors that the model can understand, specifically including: First, in the contrastive language-image preprocessing training CLIP, the text encoder transforms the abstract description of the user-input text into a high-dimensional semantic vector. , In the field of pattern image processing, textual semantic feature vectors aligned with the visual feature space are constructed by analyzing technical terms and cultural symbolic metaphors; an image encoder is then used to encode pattern images into visual feature vectors of the same dimension. , Image visual feature vectors representing pattern color contrast and pattern structural complexity; Then, the cosine similarity between the text semantic feature vector and the image visual feature vector is calculated. ; Step 3.3: Apply the pattern image generation model trained in Step 2, and through the dedicated pattern image LoRA adapter node, achieve precise control and generation of pattern features; Step 3.4: Real-time Product Generation: The pattern image output from Step 3.3 is used as input. At the same time, the semantic description of product type, material properties and usage scenario is parsed through CLIP text encoder, and the product white model LoRA is loaded to maintain the geometric integrity of the product. A cross-attention mechanism is used to achieve spatial alignment between pattern features and product structure. Combined with Nunchaku's 4-bit quantization acceleration and progressive generation strategy, high-fidelity fusion of pattern and product appearance is achieved to generate product design scheme image.

2. The product design method based on generative artificial intelligence as described in claim 1, characterized in that: In step 1, the collected pattern image data is a batik pattern image with a resolution of 1024×1024.

3. The product design method based on generative artificial intelligence as described in claim 1, characterized in that: In step 1, the preprocessing includes brightness adjustment, color adjustment, and normalization.

4. The product design method based on generative artificial intelligence as described in claim 1, characterized in that: In step 2, the low-rank adaptive method is trained using the Kohya_ss toolbox for LoRA training. During training, the parameter update matrix of the pre-trained language model is obtained by decomposing the weight matrix. Designed as a low-rank matrix and The product of the two weights preserves the basic weight matrix. Frozen state, only through low-rank matrix and To implement parameter updates, it is represented as follows: ； ； Among them, the original weight matrix , Represents the original weight matrix the number of rows, Represents the original weight matrix The number of columns.

5. The product design method based on generative artificial intelligence as described in claim 1, characterized in that: In step 2, the low-rank matrix By reconstructing the low-rank subspace features to the dimension of the basic weight matrix, we can deeply analyze the knowledge of pattern craftsmanship, understand the metaphorical system and cultural connotations of the patterns, and thus... is the rank parameter, representing the dimension of the low-rank subspace.

6. The product design method based on generative artificial intelligence as described in claim 5, characterized in that: The rank parameter ,satisfy Less than or equal to and The minimum value, i.e. The value can be 1, 2, 4 or 8.

7. The product design method based on generative artificial intelligence as described in claim 1, characterized in that: In step 2, the low-rank matrix The pattern image features input from the dataset are mapped to a low-rank subspace, thereby capturing the shallow set topological features of the patterns.

8. The product design method based on generative artificial intelligence as described in claim 1, characterized in that: In step 2, the low-rank matrix and Initialization: Low-rank matrix Using a mean of 0 and a standard deviation of , Gaussian distribution initialization This allows for the introduction of controllable randomness to explore potential combination patterns of pattern features; low-rank matrix Then initialize it as a zero matrix. This ensures the initial training phase This preserves the original generation and understanding capabilities of the model, avoiding damage to the general representation capabilities of the trained model.

9. A product design method based on generative artificial intelligence as described in claim 1, characterized in that: In step 2, the low-rank matrix and The projection matrix of query, key, and value in the cross-attention module U-Net applied to the image generation model FLUX 1.

0. , , In the formula, it is: ； ； ； in, It is a query matrix. It is a key matrix. It is a value matrix. These are the original weights of the Query projection matrix in the pre-trained model, which map each token in the input sequence to the query space and are used to calculate the attention score. These are the original weights of the Key projection matrix, which map each token of the input sequence to the key space for matching in queries; These are the original weights of the Value projection matrix in the pre-trained model, which map each token of the input sequence to the value space and are used to generate the attention-weighted context representation. and It is a low-rank matrix and The parameter matrix corresponding to the Query projection matrix; and It is a low-rank matrix and The parameter matrix corresponding to the Key projection matrix; and It is a low-rank matrix and The parameter matrix corresponding to the Value projection matrix.

10. A product design method based on generative artificial intelligence as described in claim 1, characterized in that: In step 3.2, the cosine similarity between the text semantic feature vector and the image visual feature vector is calculated. , represented as: ； in, , The closer to 1, the more likely the two vectors are in the same direction; the closer to -1, the more likely they are in opposite directions. When , it indicates orthogonality, meaning there is no correlation.

Images