Paper-cut image generation method and system based on Stable Diffusion, medium and equipment

Abstract

The invention discloses a paper-cut image generation method and system based on Stable Diffusion, a medium and equipment. The system comprises a data acquisition module, a cutting and marking module, a model training module, a model deployment module, a screen module and an adjustment module which are in communication connection. The method comprises the steps of obtaining a paper-cut image data set; cutting and marking each picture in the paper-cut pattern data set in sequence; training a text and graph generation network model by using the cut and marked paper-cut picture data set to obtain a trained paper-cut image generation network model; the trained paper-cut image generation network model is deployed to ConfyUI workflow, and the ConfyUI workflow is utilized to generate the target paper-cut image according to the text description, so that the paper-cut design efficiency can be improved, the design threshold is reduced, and the inheritance and innovative development of non-abandoned paper-cut art are promoted.

Description

Technical Field

[0001] The present invention relates to a paper-cut image generation method, system, medium and device, and more specifically, to a paper-cut image generation method, system, medium and device based on Stable Diffusion, belonging to the field of artificial intelligence technology. Background Art

[0002] Intangible cultural heritage paper-cutting art is a traditional Chinese folk art with a long history. With its unique artistic style and profound cultural connotation, it has become an important part of China's intangible cultural heritage. The history of paper-cutting art can be traced back to the Northern Dynasties, which is more than 1,500 years ago. At that time, people had mastered quite sophisticated paper-cutting skills. After the Sui and Tang Dynasties, paper-cutting art became increasingly prosperous. In the Song Dynasty, paper-cutting began to become popular, and there appeared industry artists who took paper-cutting as their profession. The Ming and Qing Dynasties were the peak period of paper-cutting. The paper-cutting handicraft art matured and was widely used in folk decoration. Its artistic characteristics: 1) Simple and bright: paper-cutting works often express profound connotations and rich imagination through simple lines and blocks; 2) Exaggeration and deformation: paper-cutting artists often exaggerate and deform the real image to make the work more decorative and artistic; 3) Symbolism and meaning: paper-cutting works often express people's good wishes and blessings for life through symbols and allegorical techniques.

[0003] At present, the inheritance and development of intangible cultural heritage paper-cutting art is becoming increasingly severe. The team of paper-cutting talents is gradually declining. Paper-cutting patterns lack creativity and artistic charm from the construction of formal sense to the presentation of cultural content, and cannot convey the cultural value of the new era. Paper-cutting design is in the manual stage, with a long design cycle and low efficiency. If the development of paper-cutting art still follows the traditional path and direct introduction paradigm, it will undoubtedly lead to its loss of development momentum in the era. Therefore, with the development of artificial intelligence technology, how to use modern scientific and technological means to design paper-cutting that meets the personalized emotional needs of consumers, so as to better connect intangible cultural heritage paper-cutting art with modern life and further stimulate the vitality of intangible cultural heritage paper-cutting culture has become a problem that needs to be solved urgently. Summary of the invention

[0004] In order to solve the above-mentioned problems of the prior art, the present invention provides a paper-cutting image generation method, system, medium and equipment based on Stable Diffusion, which has the technical characteristics of realizing the automation and intelligence of the paper-cutting design process, improving the design efficiency, lowering the threshold of paper-cutting design, promoting the inheritance and innovative transformation of intangible cultural heritage paper-cutting art in modern society, and promoting the development of the paper-cutting cultural industry and the prosperity of social culture.

[0005] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0006] The present invention discloses a paper-cut image generation system based on Stable Diffusion, comprising a data acquisition module connected by communication, a cutting and marking module, a model training module, a model deployment module, a screen module, and an adjustment module;

[0007] A data acquisition module, used to acquire a paper-cut image dataset;

[0008] A cutting and marking module, which is in communication with the data acquisition module and is used to cut and mark each image in the paper-cut image data set in turn;

[0009] A model training module is connected to the cutting and marking module for training the text-image generation network model using the cut and marked paper-cut image data set to obtain a trained network model;

[0010] A model deployment module, which is in communication with the training module and is used to deploy the trained network model into the ConfyUI workflow and generate a target paper-cut image using the ConfyUI workflow;

[0011] Screen module, which displays content in real time through a touch operation interface;

[0012] An adjustment module is communicatively connected to the model deployment module, and is used to adjust the text description if it is detected that the target paper-cut image does not meet the requirements, and use the ConfyUI workflow to generate the target paper-cut image according to the adjusted text description.

[0013] The data acquisition module refers to collecting a large number of paper-cut images through the Internet, books, field research and other methods, and then screening the collected paper-cut images. The image screening follows the principles of completeness and clarity, and removes repeated paper-cuts, partially displayed paper-cuts, paper-cuts that are incomplete due to age, and paper-cuts with insufficient clarity in the collected images, and ensures that the training set has rich and diverse paper-cut types.

[0014] The cropping and labeling module refers to the unified processing of each image in the paper-cut image dataset with the help of the image preprocessing function of StableDiffusion, all images are cropped to preset pixels, and then, the BLIP model is used to assign paper-cut feature labels to each image and the labels are adjusted and optimized in a manual manner.

[0015] The training module refers to the process of training the Stable Diffusion model with the cropped and labeled paper-cut image dataset, in which the LoRA model is used to fine-tune the Stable Diffusion model to obtain a trained paper-cut image generation network model.

[0016] The model deployment module is connected in communication with the training module, and is used to deploy the trained network model into the ConfyUI workflow, and use the ConfyUI workflow to generate the target paper-cut image; it also includes an adjustment module connected in communication with the deployment module, and the adjustment module is used to adjust the text description if it detects that the target paper-cut image does not meet the requirements, and use the ConfyUI workflow to generate the target paper-cut image according to the adjusted text description.

[0017] The present invention provides a paper-cut image generation method based on Stable Diffusion, the method comprising the following steps:

[0018] Step S1: Obtain a paper-cut image dataset;

[0019] Step S2: cutting and marking each image in the paper-cut pattern data set in turn;

[0020] Step S3: using the cropped and labeled paper-cut image data set to train the text-image generation network model to obtain a trained paper-cut image generation network model;

[0021] Step S4: deploy the trained paper-cut image generation network model into the ConfyUI workflow, and use the ConfyUI workflow to generate a target paper-cut image according to the text description.

[0022] Preferably, step S1 specifically includes the following steps:

[0023] It refers to collecting a large number of paper-cut images (such as the Internet, books, and field research), and then screening the collected paper-cut images. The screening of images follows the principles of completeness and clarity, and removes repeated paper-cuts, partially displayed paper-cuts, and paper-cut images that are incomplete and / or lack clarity due to their age, and ensures that the types of paper-cuts in the training set are diverse.

[0024] Preferably, step S2 specifically includes the following steps:

[0025] First, each image in the paper-cut image dataset is cropped to a preset pixel (such as 512*512 / 512*768, etc.), and then the image is cropped using an automated script or image processing software to retain the main body of the paper-cut pattern;

[0026] Secondly, assign detailed labels to each picture (such as the color of the paper-cut, what design elements the paper-cut pattern includes, and describe these design elements in detail as well as some paper-cutting techniques, style and other labels);

[0027] Finally, set the name of each paper-cut image and the name of the label text to correspond one to one.

[0028] Preferably, step S3 uses the method of applying the LoRA model to fine-tune the Stable Diffusion generation process to realize the automatic generation of the paper-cut image, which specifically includes the following steps:

[0029] First, train the paper-cut LoRA fine-tuning network. The process of paper-cut LoRA model training is to input the training set, set parameters, and output the model. The core parameters that determine the training effect of the paper-cut LoRA model are the training model Pretrained_model, training related parameters Repeat, Epoch, Batch_size, learning rate and optimizer unte_lr, Text_encoder_lr, Optimizer_type, and network Network_Dim;

[0030] According to the overall style characteristics of paper-cutting, the training model Pretrained_model uses the SD-F.1-dev-fp8 large model. If the training parameters of the plane model are too large, it is easy to cause overfitting, and if they are too small, it will not be able to complete the learning of the paper-cutting style characteristics. After repeated tests, the parameters finally selected for the paper-cutting LoRA model are Repeat 80, Epoch 6, Batch_size1, and the learning rate, optimizer, and network settings are unte_lr: 1e-4, text_encoder_lr: 1e-5, Optimizer_type: AdamW8, Network_Dim: 128, Network_Alpha: 64;

[0031] During the training process, the paper-cutting LoRA model will input the prompt words and pictures as conditions into the paper-cutting LoRA model. The features in the trained paper-cutting LoRA model are represented by all the features in the picture minus the features described by the prompt words, so as to use text to represent other elements that are not related to the paper-cutting features as much as possible;

[0032] Secondly, call the trained paper-cut LoRA model to fine-tune the image generation process of the Stable Diffusion model to verify the expression effect of the paper-cut feature factor; the stylized model of the Stable Diffusion model uses F.1-dev-fp8, the prompt word relevance CFG Scale is 12, the model trigger word paper-cut is input in the prompt word, the generated image resolution is 512*512, the generated image sampling step is set to 80, the prompt word remains unchanged, and the call weight of the paper-cut LoRA model is set to 0-1 to generate paper-cut images respectively; according to the generation results, the paper-cut LoRA model weight 0.9 is selected to achieve better expression of the paper-cut feature factor;

[0033] Then, keeping the parameters of the Stable Diffusion model unchanged, the weight of the paper-cut LoRA model is set to 0.9, and 10 or more groups of text prompt words describing different paper-cut features are used to generate paper-cut images to verify the generalization effect of the model; through model verification, the stylized model of the Stable Diffusion model uses F.1-dev-fp8, the weight of the paper-cut LoRA model is set to 0.9, and combined with the prompt words containing the trigger word paper-cut, it can generate new paper-cut images with corresponding prompt word features, thereby realizing the automatic generation of paper-cut images.

[0034] Preferably, step S4 specifically includes the following steps:

[0035] Deploy the Stable Diffusion model after fine-tuning the paper-cut LoRA model into the ConfyUI workflow to design an intuitive operation interface, where users can submit text prompts through the input box, including trigger words, paper-cut colors, and design preferences;

[0036] After the user submits the request, the system automatically generates a paper-cut image based on the text and options provided by the user and displays it instantly on ConfyUI. The system also allows the user to save or further modify and adjust the text description until a satisfactory paper-cut image is obtained.

[0037] The present invention provides a storage medium suitable for a paper-cut image generation method, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the paper-cut image generation method based on Stable Diffusion as claimed in claim 2 is implemented.

[0038] The present invention provides a device suitable for a paper-cut image generation method, comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor. When the processor runs the computer program, the paper-cut image generation method based on Stable Diffusion as described in claim 8 is implemented.

[0039] Beneficial effects: The present invention innovates the paper-cutting design process, realizes the automation and intelligence of the paper-cutting design process, improves the design efficiency, and also lowers the threshold of paper-cutting design through technical means. Even personnel without relevant experience or insufficient experience can design paper-cutting patterns that meet diversified personalized needs with the assistance of the model of the present invention, which promotes the inheritance and innovative transformation of intangible cultural heritage paper-cutting art in modern society, and is conducive to promoting the development of the paper-cutting cultural industry and the prosperity of social culture.

[0040] It has super flexibility and personalized customization capabilities; the introduction of ConfyUI workflow allows users to easily adjust design preferences on an intuitive interface, making design iteration efficient and intuitive. Users can instantly adjust design plans according to market demand or personal preferences without relying on complex graphic editing software or manual modification, which greatly shortens the paper-cutting design cycle;

[0041] By optimizing the Stable Diffusion model and adapting it to LoRA lightweight, the demand for hardware resources has been effectively reduced, making high-quality pattern generation no longer exclusive to high-end computing devices. This allows a wider user group to access the system and enjoy personalized paper-cutting design services, thus promoting the inheritance and development of the intangible cultural heritage of paper-cutting art. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0043] Figure 2 It is a schematic diagram of the system framework structure of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described below in conjunction with the accompanying drawings, but the present invention is not limited to the following embodiments.

[0045] like Figure 1-2 The present invention shows a specific embodiment of a paper-cut image generation method, system, medium and device based on Stable Diffusion, wherein a paper-cut image generation system based on Stable Diffusion includes a data acquisition module connected by communication, a cutting and marking module, a model training module, a model deployment module, a screen module and an adjustment module;

[0046] The data acquisition module is used to obtain a paper-cut image dataset; specifically: the data acquisition module collects a large number of paper-cut images through the Internet, books, field research, etc., and then screens the collected paper-cut images. The image screening follows the principles of completeness and clarity, and removes repeated paper-cuts, partially displayed paper-cuts, paper-cuts that are incomplete due to age, and paper-cuts with insufficient clarity in the collected images, and ensures that the training set has rich and diverse paper-cut types.

[0047] The cropping and labeling module is connected to the data acquisition module for cropping and labeling each image in the paper-cut image data set in turn. Specifically, the cropping and labeling module processes each image in the paper-cut image data set in a unified manner with the help of the image preprocessing function of Stable Diffusion, and all images are cropped to preset pixels. Then, the BLIP model is used to assign a paper-cut feature label to each image and the label is adjusted and optimized in a manual manner.

[0048] The model training module is in communication connection with the cutting and marking module, and is used to train the text-image generation network model using the paper-cut image data set after cutting and marking, and obtain the trained network model; specifically: in the process of training the Stable Diffusion model using the paper-cut image data set after cutting and marking, the model training module uses the LoRA model to fine-tune the Stable Diffusion model, and obtains the trained paper-cut image generation network model;

[0049] A model deployment module, which is in communication with the training module, is used to deploy the trained network model into the ConfyUI workflow, and use the ConfyUI workflow to generate a target paper-cut image; specifically: the model deployment module is used to deploy the trained network model into the ConfyUI workflow, and use the ConfyUI workflow to generate a target paper-cut image;

[0050] The adjustment module is used to adjust the text description if it is detected that the target paper-cut image does not meet the requirements, and use the ConfyUI workflow to generate the target paper-cut image according to the adjusted text description. The screen module displays the content in real time through a touch operation interface.

[0051] The present invention provides a paper-cut image generation method based on Stable Diffusion, the method comprising the following steps:

[0052] Step S1: Obtain a paper-cut image dataset; specifically, the following steps are included: a large number of paper-cut images are collected (such as from the Internet, books, and field research), and then the collected paper-cut images are screened. The image screening follows the principles of completeness and clarity, and duplicate paper-cuts, partially displayed paper-cuts, and paper-cut images that are incomplete and / or lack clarity due to their age are eliminated from the collected images, and the types of paper-cuts in the training set are ensured to be diverse.

[0053] Step S2: Crop and label each image in the paper-cut pattern data set in turn; specifically, the following steps are included: first, crop each image in the paper-cut image data set to preset pixels (such as 512*512 / 512*768, etc.), and then use an automated script or image processing software to crop the image to retain the main body of the paper-cut pattern; second, assign detailed labels to each image (such as the color of the paper-cut, which design elements the paper-cut pattern includes, and describe these design elements and some paper-cutting techniques, style and other labels in detail); finally, set the name of each paper-cut image and the name of the label text to correspond one to one.

[0054] Step S3: Use the cropped and labeled paper-cut image dataset to train the text-image generation network model to obtain a trained paper-cut image generation network model; use the LoRA model to fine-tune the Stable Diffusion generation process to achieve automatic generation of paper-cut images, specifically including the following steps:

[0055] First, train the paper-cut LoRA fine-tuning network. The process of paper-cut LoRA model training is to input the training set, set parameters, and output the model. The core parameters that determine the training effect of the paper-cut LoRA model are the training model Pretrained_model, training related parameters Repeat, Epoch, Batch_size, learning rate and optimizer unte_lr, Text_encoder_lr, Optimizer_type, and network Network_Dim;

[0056] According to the overall style characteristics of paper-cutting, the training model Pretrained_model uses the SD-F.1-dev-fp8 large model. If the training parameters of the plane model are too large, it is easy to cause overfitting, and if they are too small, it will not be able to complete the learning of the paper-cutting style characteristics. After repeated tests, the parameters finally selected for the paper-cutting LoRA model are Repeat 80, Epoch 6, Batch_size1, and the learning rate, optimizer, and network settings are unte_lr: 1e-4, text_encoder_lr: 1e-5, Optimizer_type: AdamW8, Network_Dim: 128, Network_Alpha: 64;

[0057] During the training process, the paper-cutting LoRA model will input the prompt words and pictures as conditions into the paper-cutting LoRA model. The features in the trained paper-cutting LoRA model are represented by all the features in the picture minus the features described by the prompt words, so as to use text to represent other elements that are not related to the paper-cutting features as much as possible;

[0058] Secondly, call the trained paper-cut LoRA model to fine-tune the image generation process of the Stable Diffusion model to verify the expression effect of the paper-cut feature factor; the stylized model of the Stable Diffusion model uses F.1-dev-fp8, the prompt word relevance CFG Scale is 12, the model trigger word paper-cut is input in the prompt word, the generated image resolution is 512*512, the generated image sampling step is set to 80, the prompt word remains unchanged, and the call weight of the paper-cut LoRA model is set to 0-1 to generate paper-cut images respectively; according to the generation results, the paper-cut LoRA model weight 0.9 is selected to achieve better expression of the paper-cut feature factor;

[0059] Then, keeping the parameters of the Stable Diffusion model unchanged, the weight of the paper-cut LoRA model is set to 0.9, and 10 or more groups of text prompt words describing different paper-cut features are used to generate paper-cut images to verify the generalization effect of the model; through model verification, the stylized model of the Stable Diffusion model uses F.1-dev-fp8, the weight of the paper-cut LoRA model is set to 0.9, and combined with the prompt words containing the trigger word paper-cut, it can generate new paper-cut images with corresponding prompt word features, thereby realizing the automatic generation of paper-cut images.

[0060] Step S4: deploy the trained paper-cut image generation network model to the ConfyUI workflow, and use the ConfyUI workflow to generate the target paper-cut image according to the text description; specifically including the following steps: deploy the Stable Diffusion model after fine-tuning the paper-cut LoRA model to the ConfyUI workflow to design an intuitive operation interface, and the user can submit text prompts through the input box, including trigger words, paper-cut colors, and design preferences; after the user submits the request, the system automatically generates a paper-cut image based on the text and options provided by the user and displays it instantly on ConfyUI, and allows the user to save or further modify and adjust the text description until a satisfactory paper-cut image is obtained.

[0061] The present invention discloses a storage medium suitable for a paper-cut image generation method. The storage medium stores a computer program. When the computer program is executed by a processor, the paper-cut image generation method based on Stable Diffusion is implemented.

[0062] Storage media, as the physical basis of data storage, are of various types, each with its own characteristics, and are suitable for different application scenarios. In this application, the storage medium can be one or more of a disk storage medium, an optical disk storage medium, a flash storage medium, and a tape storage medium. For example: LTO (Linear Tape-Open) tape: an open tape format that supports high-capacity, high-speed data storage. For example, IBM's LTO-8 tape, with a single-disk capacity of up to 30TB, is very suitable for cold data storage and long-term backup.

[0063] The present invention discloses a device suitable for a paper-cut image generation method, comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor. When the processor runs the computer program, the paper-cut image generation method based on Stable Diffusion is implemented.

[0064] The processor (CPU, Central Processing Unit) is responsible for executing instructions in the program, processing data, and controlling other components of the computer. This application does not limit the specifications and models of the processor.

[0065] The processor can be:

[0066] Intel Core i5-11400F: Supports advanced instruction sets such as SSE, AVX2, and Intel's Hyper-Threading technology, which improves the processor's multi-threaded processing capabilities.

[0067] Intel Core i7-11700K: Intel's Core series processor, i7 represents the performance level, 11700K represents the model number, and "K" indicates that the processor supports overclocking.

[0068] AMD Ryzen 9 5950X: has 16 cores and 32 threads.

[0069] Intel Core i9-12900K: The base frequency is 3.2GHz and the maximum turbo frequency can reach 5.2GHz.

[0070] AMD Ryzen 7 5800X: Equipped with 32MB of L3 cache, it provides the processor with ample data storage space and improves data access efficiency.

[0071] Finally, it should be noted that the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by ordinary technicians in this field should be considered as the protection scope of the present invention.

Claims

1. A paper-cut image generation system based on Stable Diffusion, characterized by: It includes a data acquisition module for communication connection, a cropping and marking module, a model training module, a model deployment module, a screen module, and an adjustment module; A data acquisition module, used to acquire a paper-cut image dataset; A cutting and marking module, which is in communication with the data acquisition module and is used to cut and mark each image in the paper-cut image data set in turn; A model training module is connected to the cutting and marking module for training the text-image generation network model using the cut and marked paper-cut image data set to obtain a trained network model; A model deployment module, which is in communication with the model training module, is used to deploy the trained network model into the ConfyUI workflow, and generate a target paper-cut image using the ConfyUI workflow; Screen module, which displays content in real time through a touch operation interface; An adjustment module is communicatively connected to the model deployment module, and is used to adjust the text description if it is detected that the target paper-cut image does not meet the requirements, and use the ConfyUI workflow to generate the target paper-cut image according to the adjusted text description.

2. A paper-cut image generation method based on Stable Diffusion, characterized in that: Using the system as claimed in claim 1, the method comprises the following steps: Step S1: Obtain a paper-cut image dataset; Step S2: cutting and marking each image in the paper-cut pattern data set in turn; Step S3: using the cropped and labeled paper-cut image data set to train the text-image generation network model to obtain a trained paper-cut image generation network model; Step S4: deploy the trained paper-cut image generation network model into the ConfyUI workflow, and use the ConfyUI workflow to generate a target paper-cut image according to the text description.

3. The method for generating paper-cut images based on Stable Diffusion according to claim 2, characterized in that: Step S1 specifically includes the following steps: It means collecting a large number of paper-cut images and then screening the collected paper-cut images. The screening of images follows the principles of completeness and clarity, and removes repeated paper-cuts, partially displayed paper-cuts, incomplete and / or unclear paper-cut images in the collected images, and ensures that the types of paper-cuts in the training set are diverse.

4. The method for generating paper-cut images based on Stable Diffusion according to claim 2, characterized in that: Step S2 specifically includes the following steps: First, each image in the paper-cut image dataset is cropped to a preset pixel, and then the image is cropped using an automated script or image processing software to retain the main body of the paper-cut pattern; Second, assign detailed labels to each image; Finally, set the name of each paper-cut image and the name of the label text to correspond one to one.

5. The method for generating paper-cut images based on Stable Diffusion according to claim 2, characterized in that: Step S3 uses the LoRA model to fine-tune the Stable Diffusion generation process to achieve automatic generation of paper-cut images. The following steps are involved: First, train the paper-cut LoRA fine-tuning network. The process of paper-cut LoRA model training is to input the training set, set parameters, and output the model. The core parameters that determine the training effect of the paper-cut LoRA model are the training model Pretrained_model, training related parameters Repeat, Epoch, Batch_size, learning rate and optimizer unte_lr, Text_encoder_lr, Optimizer_type, and network Network_Dim; The training model Pretrained_model uses the SD-F.1-dev-fp8 large model. After repeated tests, the parameters selected for the Papercut LoRA model are Repeat 80, Epoch 6, Batch_size1. At the same time, the learning rate, optimizer, and network settings are unte_lr: 1e-4, text_encoder_lr: 1e-5, Optimizer_type: AdamW8, Network_Dim: 128, Network_Alpha: 64; During the training process, the paper-cutting LoRA model will input the prompt words and pictures as conditions into the paper-cutting LoRA model. The features in the trained paper-cutting LoRA model are represented by all the features in the picture minus the features described by the prompt words, so as to use text to represent other elements that are not related to the paper-cutting features as much as possible; Secondly, call the trained paper-cut LoRA model to fine-tune the image generation process of the Stable Diffusion model to verify the expression effect of the paper-cut feature factor; the stylized model of the Stable Diffusion model uses F.1-dev-fp8, the prompt word relevance CFG Scale is 12, the model trigger word paper-cut is input in the prompt word, the generated image resolution is 512*512, the generated image sampling step is set to 80, the prompt word remains unchanged, and the call weight of the paper-cut LoRA model is set to 0-1 to generate paper-cut images respectively; according to the generation results, the paper-cut LoRA model weight 0.9 is selected to achieve better expression of the paper-cut feature factor; Then, keeping the parameters of the Stable Diffusion model unchanged, the weight of the paper-cut LoRA model is set to 0.9, and 10 groups of text prompt words describing different paper-cut features are used to generate paper-cut images to verify the generalization effect of the model; through the verification of the model, the stylized model of the Stable Diffusion model uses F.1-dev-fp8, the weight of the paper-cut LoRA model is set to 0.9, and combined with the prompt words containing the trigger word paper-cut, it can generate new paper-cut images with corresponding prompt word features, thereby realizing the automatic generation of paper-cut images.

6. The method for generating paper-cut images based on Stable Diffusion according to claim 2, characterized in that: Step S4 specifically includes the following steps: Deploy the Stable Diffusion model after fine-tuning the paper-cut LoRA model into the ConfyUI workflow to design an intuitive operation interface, where users can submit text prompts through the input box, including trigger words, paper-cut colors, and design preferences; After the user submits the request, the system automatically generates a paper-cut image based on the text and options provided by the user and displays it instantly on ConfyUI. The system also allows the user to save or further modify and adjust the text description until a satisfactory paper-cut image is obtained.

7. A storage medium suitable for a paper-cut image generation method, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the paper-cut image generation method based on Stable Diffusion as claimed in claim 2 is implemented.

8. A device suitable for a paper-cut image generation method, characterized in that: It includes a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor. When the processor runs the computer program, the paper-cut image generation method based on Stable Diffusion as described in claim 7 is implemented.

Images