Commodity creative design method and system supporting structured text input
By constructing product image data sets and training title generation models, and using images and text encoders to generate structured text encoding, the problem that personalized needs in traditional design methods is difficult to meet, and efficient personalized creative design is achieved.
Patent Information
- Application Number
- CN202311851602.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-29
- Publication Date
- 2025-08-08
AI Technical Summary
Traditional creative product design methods are difficult to meet users' personalized needs. The designer's explanation of the user's oral or verbal descriptions is vague and subjective, and the creative design process is time-consuming and costly.
The product image data set is constructed and preprocessed, the title generation model is trained, the structured text encoding is generated using the image and text encoder, and the product image that conforms to the user's description is generated through the diffusion model.
Improve the accuracy and efficiency of creative understanding, generate more diverse and personalized creative designs, and reduce design costs and time.
Smart Images
Figure CN120451333A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of commodity design, and in particular to a commodity creative design method and system supporting structured text input. Background Art
[0002] With technological advancements and the ubiquity of social media, users are increasingly demanding personalized and customized products. Traditional product creative design methods may struggle to meet the individual needs of diverse users. The traditional creative design process typically requires designers to interpret and understand users' verbal or written descriptions, but natural language expressions can be ambiguous, cryptic, and subjective. Furthermore, different users may have significant differences in their needs and preferences for product creativity. Traditional creative design methods may be limited by designers' experience and creativity, making them incapable of meeting the demand for large-scale personalized creativity. Summary of the Invention
[0003] The present invention provides a product creative design method and system supporting structured text input to overcome the above technical problems.
[0004] A product creative design method supporting structured text input, including:
[0005] S1. Obtain product images of different categories and styles and build a product image dataset.
[0006] S2. Preprocess the product images in the product image dataset, including cropping the product images, filling blank areas in the product images, and adjusting all product images to the same resolution.
[0007] S3. Build a title generation model. Obtain a training set from the preprocessed product image dataset. For each product image in the training set, manually generate a product description. Train the title generation model based on the training set and the description of each product image in the training set.
[0008] S4. Generate a description of each product image in the product image dataset based on the trained title generation model, input the product image into the image encoder to obtain the image code, and input the description of the product image into the text encoder to obtain the text code.
[0009] S5. Construct a diffusion model and train it based on text encoding and image encoding.
[0010] S6. During the testing phase, structured text is obtained, the structured text is input into a large language model for optimization, the optimized structured text is input into a text encoder to obtain a structured text encoding, the structured text encoding is input into a trained diffusion model to obtain a denoised image encoding, the denoised image encoding is input into an image decoder to obtain a product image corresponding to the structured text.
[0011] Preferably, the text encoder is a Transformer encoder.
[0012] Preferably, the image encoder is a diffusion model.
[0013] Preferably, the diffusion model is a deep neural network algorithm model.
[0014] A product creative design system includes a data set acquisition and processing module, a title generation model construction module, an encoding module, a diffusion model construction module, and a product image generation module.
[0015] The dataset acquisition and processing module is used to acquire product images of different categories and styles and construct a product image dataset. It also preprocesses the product images in the product image dataset, including cropping product images, filling blank areas in product images, and adjusting all product images to the same resolution.
[0016] The title generation model construction module is used to construct and train the title generation model. The construction and training of the title generation model includes obtaining a training set from the preprocessed product image dataset, manually generating a product description for each product image in the training set, and training the title generation model based on the training set and the description of each product image in the training set.
[0017] The encoding module is used to encode the description and product image. The encoding of the description and product image includes generating a description of each product image in the product image dataset based on the trained title generation model, inputting the product image into the image encoder to obtain the image code, and inputting the description of the product image into the text encoder to obtain the text code.
[0018] The diffusion model construction module is used to build a diffusion model and train the diffusion model based on text encoding and image encoding.
[0019] The product image generation module is used to obtain the product image corresponding to the structured text. The obtaining of the product image corresponding to the structured text includes obtaining the structured text, inputting the structured text into a large language model for optimization, inputting the optimized structured text into a text encoder to obtain the structured text code, inputting the structured text code into a trained diffusion model and obtaining a denoised image code, inputting the denoised image code into an image decoder, and obtaining the product image corresponding to the structured text.
[0020] The present invention provides a product creative design method and system that supports structured text input. The present invention uses natural language processing technology to convert natural language text into machine-readable structured data, thereby reducing the understanding burden on designers and improving the accuracy and efficiency of creative understanding; the method that supports structured text input can generate more diverse and personalized creative designs based on the structured description provided by the user, thereby enhancing the differentiated competitive advantage of the product; the present invention uses artificial intelligence algorithms to accelerate the creative design cycle and reduce costs. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0022] Figure 1 It is a flow chart of the method of the present invention; DETAILED DESCRIPTION
[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0024] Figure 1 It is a flow chart of the method of the present invention, as shown in Figure 1 As shown, the method of this embodiment may include:
[0025] S1. Obtain product images of different categories and styles and build a product image dataset.
[0026] S2. Preprocess the product images in the product image dataset, including cropping the product images, filling blank areas in the product images, and adjusting all product images to the same resolution.
[0027] S3. Build a title generation model. Obtain a training set from the preprocessed product image dataset. For each product image in the training set, manually generate a product description. Train the title generation model based on the training set and the description of each product image in the training set.
[0028] S4. Generate a description of each product image in the product image dataset based on the trained title generation model, input the product image into the image encoder to obtain the image code, and input the description of the product image into the text encoder to obtain the text code.
[0029] S5. Construct a diffusion model and train it based on text encoding and image encoding.
[0030] S6. During the testing phase, structured text is obtained, the structured text is input into a large language model for optimization, the optimized structured text is input into a text encoder to obtain a structured text encoding, the structured text encoding is input into a trained diffusion model to obtain a denoised image encoding, the denoised image encoding is input into an image decoder to obtain a product image corresponding to the structured text.
[0031] Based on the above scheme, the present invention provides a product creative design method and system that supports structured text input. The present invention uses natural language processing technology to convert natural language text into machine-readable structured data, thereby reducing the understanding burden on designers and improving the accuracy and efficiency of creative understanding; the method that supports structured text input can generate more diverse and personalized creative designs based on the structured description provided by the user, thereby enhancing the differentiated competitive advantage of the product; the present invention uses artificial intelligence algorithms to accelerate the creative design cycle and reduce costs.
[0032] Specifically, this embodiment provides the details of the product creative design method, including:
[0033] S1. Obtain product images of different categories and styles and build a product image dataset. Specifically, in order to cope with product application scenarios of different styles or designs, it is necessary to collect product images of various categories and styles as a training set for the model. The product image dataset can be obtained through crawler algorithms or other means.
[0034] S2. Preprocess the product images in the product image dataset. The preprocessing includes cropping the product images, filling the blank areas in the product images, and adjusting all the product images to the same resolution. Specifically, the collected original images have different sizes and resolutions, so the original images need to be preprocessed. The preprocessing steps include cropping and blank filling to ensure that a set of images of consistent size is obtained.
[0035] S3. Build a title generation model. Obtain a training set from the preprocessed product image dataset. For each product image in the training set, manually generate a product description. Train the title generation model based on the training set and the description of each product image in the training set. The title generation model is the blip2 model.
[0036] Specifically, since the collected data only contains the image itself without any text description, it is necessary to generate a caption for the input image to obtain a text description that matches the image content. This can help convert the image's visual features into semantic text information, providing a reference for subsequent text encoding and image generation.
[0037] S4. Generate a description of each product image in the product image dataset based on the trained title generation model, input the product image into an image encoder to obtain an image code, and input the description of the product image into a text encoder to obtain a text code. The image encoder is a diffusion model, and the text encoder is a Transformer encoder.
[0038] Specifically, since the model cannot directly recognize text, it first needs to convert the text input into tokens, then encode the tokens to obtain a text vector. The image input is compressed and encoded by the image encoder to obtain a compressed image code.
[0039] S5. Construct a diffusion model, a deep neural network algorithm model, which is trained based on the text and image encodings. Specifically, the image encoding is input into the diffusion model, with the text vector as a conditional input, to guide the diffusion of the image in the latent space, ultimately generating a denoised image encoding. In this step, the model uses the learned association between text and image to diffuse the image encoding in the latent space, thereby generating product images that meet the prompt word requirements.
[0040] S6. Obtaining structured text and testing the model established in the above steps includes: During the testing phase, obtaining structured text and inputting it into a large language model for optimization. Specifically, a large language model is an artificial intelligence model designed to understand and generate human language. They are trained on large amounts of text data and can perform a wide range of tasks, including text summarization, translation, sentiment analysis, and more. In this embodiment, the GPT-3 model is used.
[0041] The optimized structured text is input into a text encoder to obtain a structured text code. This code is then input into a trained diffusion model to obtain a denoised image code. This denoised image code is then input into an image decoder to obtain the product image corresponding to the structured text. Specifically, the denoised image code is input into an image decoder for reconstruction, generating the final product image. The image decoder can restore the compressed image code to its original form, thereby producing a product image with a specific color, style, or pattern.
[0042] A product creative design system includes a data set acquisition and processing module, a title generation model construction module, an encoding module, a diffusion model construction module, and a product image generation module.
[0043] The dataset acquisition and processing module is used to acquire product images of different categories and styles and construct a product image dataset. It also preprocesses the product images in the product image dataset, including cropping product images, filling blank areas in product images, and adjusting all product images to the same resolution.
[0044] The title generation model construction module is used to construct and train the title generation model. The construction and training of the title generation model includes obtaining a training set from the preprocessed product image dataset, manually generating a product description for each product image in the training set, and training the title generation model based on the training set and the description of each product image in the training set.
[0045] The encoding module is used to encode the description and product image. The encoding of the description and product image includes generating a description of each product image in the product image dataset based on the trained title generation model, inputting the product image into the image encoder to obtain the image code, and inputting the description of the product image into the text encoder to obtain the text code.
[0046] The diffusion model construction module is used to build a diffusion model and train the diffusion model based on text encoding and image encoding.
[0047] The product image generation module is used to obtain the product image corresponding to the structured text. The obtaining of the product image corresponding to the structured text includes obtaining the structured text, inputting the structured text into a large language model for optimization, inputting the optimized structured text into a text encoder to obtain the structured text code, inputting the structured text code into a trained diffusion model and obtaining a denoised image code, inputting the denoised image code into an image decoder, and obtaining the product image corresponding to the structured text.
[0048] The present invention mainly solves the following problems:
[0049] 1. Creative Understanding: Traditional creative design processes typically require designers to interpret and understand users' verbal or written descriptions. However, natural language expressions can be vague, ambiguous, and subjective. This invention utilizes natural language processing technology to convert natural language text into machine-readable structured data, thereby reducing the comprehension burden on designers and improving the accuracy and efficiency of creative understanding.
[0050] 2. Creative Diversity and Personalization: Different users may have significantly different needs and preferences for product creativity. Traditional creative design methods may be limited by the designer's experience and creativity, making it difficult to meet the needs of large-scale personalized creativity. Methods that support structured text input can generate more diverse and personalized creative designs based on the structured descriptions provided by users, thereby enhancing the differentiated competitive advantage of products.
[0051] 3. Creative generation and screening: Creative design is a creative process that requires a lot of time and resources to generate and screen the best solutions. This invention can use artificial intelligence algorithms to speed up the creative design cycle and reduce costs.
[0052] The beneficial effects of the product creative design method that supports structured text input include:
[0053] 1. Improved Creative Efficiency: Structured text input makes the creative design process more efficient. Designers can express their ideas in a structured manner, reducing errors in understanding and communication. This helps speed up the generation and realization of ideas.
[0054] 2. Personalized design: Structured text input can better capture users' personalized needs. By analyzing the structured descriptions of user input, the design system can more accurately understand the user's likes and preferences, thereby generating creative designs that better meet user needs.
[0055] 3. Improve design accuracy: Structured text input helps reduce ambiguity and ambiguous descriptions. Designers and systems can understand design requirements more clearly, thus avoiding misunderstandings and incorrect design directions.
[0056] 4. Optimize user experience: Structured text input can better capture user emotions and feelings. By analyzing user emotional descriptions, the design system can create product designs that better meet users' emotional and psychological needs, thereby improving the user experience.
[0057] 5. Enhanced design feedback and iteration: Structured text input makes feedback between users and designers more direct and clear. This helps to timely adjust and improve design solutions and continuously optimize product creativity.
[0058] 6. Improve cross-team collaboration efficiency: In multi-person design projects, structured text input can better facilitate cross-team collaboration and communication. Different teams can understand the design goals together based on the structured description and work together better.
[0059] In summary, a product creative design method that supports structured text input can significantly improve the creative design process, increase efficiency, enhance personalization and user experience, and provide designers and businesses with more insights and business opportunities. This method helps create more attractive and competitive products, meet user needs, and promote sustainable business development.
[0060] Overall beneficial effects:
[0061] The present invention provides a product creative design method and system that supports structured text input. The present invention uses natural language processing technology to convert natural language text into machine-readable structured data, thereby reducing the understanding burden on designers and improving the accuracy and efficiency of creative understanding; the method that supports structured text input can generate more diverse and personalized creative designs based on the structured description provided by the user, thereby enhancing the differentiated competitive advantage of the product; the present invention uses artificial intelligence algorithms to accelerate the creative design cycle and reduce costs.
[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A product creative design method supporting structured text input, characterized in that: include, S1. Obtain product images of different categories and styles and build a product image dataset. S2. Preprocess the product images in the product image dataset, including cropping the product images, filling blank areas in the product images, and adjusting all product images to the same resolution. S3. Build a title generation model. Obtain a training set from the preprocessed product image dataset. For each product image in the training set, manually generate a product description. Train the title generation model based on the training set and the description of each product image in the training set. S4. Generate a description of each product image in the product image dataset based on the trained title generation model, input the product image into the image encoder to obtain the image code, and input the description of the product image into the text encoder to obtain the text code. S5. Construct a diffusion model and train it based on text encoding and image encoding. S6. During the testing phase, structured text is obtained, the structured text is input into a large language model for optimization, the optimized structured text is input into a text encoder to obtain a structured text encoding, the structured text encoding is input into a trained diffusion model to obtain a denoised image encoding, the denoised image encoding is input into an image decoder to obtain a product image corresponding to the structured text.
2. A product creative design method supporting structured text input according to claim 1, characterized in that: The text encoder is a Transformer encoder.
3. The method for product creative design supporting structured text input according to claim 1, characterized in that: The image encoder is a diffusion model.
4. The method for product creative design supporting structured text input according to claim 1, characterized in that: The diffusion model is a deep neural network algorithm model.
5. A product creative design system based on the method of claim 1, characterized in that: It includes dataset acquisition and processing module, title generation model construction module, encoding module, diffusion model construction module and product image generation module. The dataset acquisition and processing module is used to acquire product images of different categories and styles and construct a product image dataset. It also preprocesses the product images in the product image dataset, including cropping product images, filling blank areas in product images, and adjusting all product images to the same resolution. The title generation model construction module is used to construct and train the title generation model. The construction and training of the title generation model includes obtaining a training set from the preprocessed product image dataset, manually generating a product description for each product image in the training set, and training the title generation model based on the training set and the description of each product image in the training set. The encoding module is used to encode the description and product image. The encoding of the description and product image includes generating a description of each product image in the product image dataset based on the trained title generation model, inputting the product image into the image encoder to obtain the image code, and inputting the description of the product image into the text encoder to obtain the text code. The diffusion model construction module is used to build a diffusion model and train the diffusion model based on text encoding and image encoding. The product image generation module is used to obtain the product image corresponding to the structured text. The obtaining of the product image corresponding to the structured text includes obtaining the structured text, inputting the structured text into a large language model for optimization, inputting the optimized structured text into a text encoder to obtain the structured text code, inputting the structured text code into a trained diffusion model and obtaining a denoised image code, inputting the denoised image code into an image decoder, and obtaining the product image corresponding to the structured text.