Web-based intelligent image creation system and implementation method

Abstract

The invention provides a Web-based intelligent image creation system and an implementation method thereof. The system comprises a Web application service module, an image creation module and an image creation module, the model service cluster comprises a Docker container deployment text-to-image generation model based on Kubernetes management, an image-to-image conversion model and a low-rank adaptive model LoRA; the distributed file storage system is used for storing a user-defined model, a training data set and a generated work and supporting multi-copy redundant backup; the integrated LoRA training module provides data set labeling, trigger word binding and training parameter optimization functions; the multi-modal image processing module has the functions of local redrawing, intelligent image expansion, super-resolution repair and style migration; and the dual-channel content security auditing module adopts BERT text filtering and YOLOv5 image recognition technologies. The technical defect that an existing AI generation tool cannot support a user-defined model and a dynamic extension service is overcome, and the whole-process closed-loop operation of a user from data labeling to model deployment is supported.

Description

Technical Field

[0001] The present invention belongs to the technical field of intersection of artificial intelligence and computer vision, and specifically relates to a Web-based intelligent image creation system and implementation method, especially supporting user-defined model training and multimodal image editing. Background Art

[0002] With the development of artificial intelligence (AI) technology, AI is increasingly being used in image generation and editing. Traditional image creation relies on specialized software (such as Photoshop and Maya), which is complex to use and has high learning costs. While existing AI generation tools (such as Midjourney and Stable Diffusion WebUI) provide basic generation capabilities, they suffer from the following drawbacks and are difficult to resolve: Model closure: It cannot support users to upload custom generated models or fine-tune adaptation layers (such as LoRA); Single function: lack of integrated image secondary editing tools (such as partial redrawing, intelligent repair, etc.); Poor service scalability: The single-machine deployment model cannot handle high-concurrency requests, and the generation efficiency drops sharply as the load increases; Content security risks: There is a lack of a dual-channel input and output review mechanism, and there is a risk of spreading illegal content.

[0003] The significance of solving the above problems is: through containerized deployment, the service response speed is increased by 40%, and thousands of concurrent tasks are supported; the user-defined model function increases the creative freedom by 300%, and supports personalized model sharing; a one-stop multi-level image processing pipeline, including local redrawing, intelligent image expansion, super-resolution restoration and style transfer functions; the built-in LoRA training module reduces the training cost of professional users by 80%, shortens the training cycle to 30 minutes, and supports users' full-process closed-loop operation from data annotation to model deployment; the dual-channel content review mechanism of text input and image output achieves a 99.8% illegal content interception rate. Summary of the Invention

[0004] The present invention provides a Web-based intelligent image creation system and implementation method to solve the problems in the prior art of insufficient flexibility in image creation and weak personalized customization capabilities.

[0005] According to one aspect of the present invention, a Web-based intelligent image creation system is provided, comprising: a Web application service module, which uses the Vue.js 3 framework to build a front-end interactive interface, implements the back-end business logic through the Spring Boot framework, and communicates between the front-end and the back-end through a RESTful API; a model service cluster, which includes a text-to-image generation model, an image-to-image conversion model, and a low-rank adaptive model LoRA deployed in Docker containers managed by Kubernetes, and supports dynamic allocation of computing resources according to load; a distributed file storage system (DFS) for storing user-defined models, training data sets, and generated works, and supports multiple copies of redundant backup; a LoRA model training module, which integrates a data set annotation tool, a training parameter configuration interface, and a model testing environment, and supports trigger word binding and dynamic adjustment of the learning rate; a multimodal image processing module, which includes a local redrawing function submodule based on an attention mechanism, an intelligent image expansion function submodule based on a generative adversarial network (GAN), and a super-resolution restoration function submodule based on a convolutional neural network (CNN); and a dual-channel content security review module, which uses natural language processing technology to filter the input text for compliance and uses computer vision technology to perform content recognition on the generated images.

[0006] Optionally, in the above-mentioned Web-based intelligent image creation system, the model service cluster includes: a heartbeat monitoring mechanism, each container node sends status information including service type, model version and resource occupancy rate to the database every 10 seconds; and a service routing module, which dynamically generates a list of available service nodes based on the heartbeat information and allocates generation tasks through a weighted polling algorithm.

[0007] Optionally, in the above-mentioned Web-based intelligent image creation system, the LoRA model training module includes: an automatic annotation unit for semantically annotating training images; a parameter optimization unit for adjusting the text encoder learning rate and the U-Net learning rate; and a trigger word binding interface for associating and mapping specific text tags with training data features.

[0008] Optionally, in the above-mentioned Web-based intelligent image creation system, the multimodal image processing module includes: an intelligent face-changing sub-module, which uses StyleGAN2 to encode and transfer facial features; a super-resolution reconstruction sub-module, which uses ESRGAN to achieve a 4x resolution improvement; and a progressive restoration algorithm, which restores texture details of old photos through a multi-scale residual network.

[0009] Optionally, in the above-mentioned Web-based intelligent image creation system, the content security review module includes: a text filtering layer, which uses the BERT-base model to detect sensitive political keywords; an image review layer, which uses YOLOv5 to identify illegal visual elements; and an audit tracking system, which records the hash values and audit logs of all generated content.

[0010] Optionally, in the above-mentioned Web-based intelligent image creation system, an integrated LoRA model training module is used to provide data set annotation, trigger word binding and training parameter optimization functions.

[0011] Optionally, in the above-mentioned Web-based intelligent image creation system, the content security review module adopts a dual-channel content security review mechanism, using BERT text filtering and YOLOv5 image recognition technology.

[0012] According to another aspect of the present invention, an implementation method of the above-mentioned Web-based intelligent image creation system is provided, including: S1. building a web application service; S2. using Kubernetes to build the docker container cluster; and S3. building a distributed file storage system and a file server, wherein the distributed file storage system is used to store various types of text-based image models, image-based image models, LoRA models, and user-created image works, and using nginx to build the file server and mount the corresponding distributed storage directory.

[0013] According to the technical solution of the present invention, the beneficial effects produced are: By integrating advanced AI technology with a user-friendly design, this invention empowers users with unprecedented creative freedom while maintaining high efficiency. It not only transcends the functional limitations of traditional image editing tools but also creates a new interactive model, enabling more people to participate in digital art creation. This represents a significant improvement for both professional designers seeking a more efficient workflow and ordinary users seeking to easily express their creativity.

[0014] In order to better understand and illustrate the concept, working principle and effect of the present invention, the present invention is described in detail below with reference to the accompanying drawings and through specific embodiments: BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the specific implementation or the description of the prior art.

[0016] Figure 1 is an architectural diagram of a Web-based intelligent image creation system according to an example embodiment.

[0017] Figure 2 It is a flowchart of the architecture construction of a Web-based intelligent image creation system according to an example embodiment.

[0018] Figure 3 A flowchart of an image creation process is provided according to an exemplary embodiment.

[0019] Figure 4 1 is a LoRA training flow chart provided according to an example embodiment. DETAILED DESCRIPTION

[0020] To make the purpose, technical methods and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and specific examples. These examples are merely illustrative and not intended to limit the present invention.

[0021] like Figure 1 As shown, the web-based intelligent image creation system of the present invention includes: a web application service module, a model service cluster, a distributed file storage system (DFS), a LoRA model training module, a multimodal image processing module, and a content security audit module, wherein: The web application service module uses the Vue.js 3 framework to build the front-end interactive interface, the Spring Boot framework to implement the back-end business logic, and the RESTful API for front-end and back-end communication. The front-end and back-end separation architecture based on Vue.js 3 and Spring Boot provides cross-platform access capabilities. Model serving cluster, including text-to-image generation models deployed in Docker containers managed by Kubernetes (i.e. Figure 1 The cultural graph model in ), image-to-image translation model ( Figure 1 The model service cluster includes a heartbeat monitoring mechanism, where each container node sends status information including service type, model version, and resource utilization to the database every 10 seconds; and a service routing module, which dynamically generates a list of available service nodes based on heartbeat information and distributes generation tasks using a weighted round-robin algorithm.

[0022] Distributed File Storage System (DFS) is used to store user-defined models, training data sets and generated works, and supports multiple copies of redundant backups; the user-defined model upload module supports distributed storage of text-based graph models, LoRA models and data sets The LoRA model training module includes: an automatic annotation unit that supports the selection of multimodal models such as CLIP and BLIP to perform semantic annotation on training images; a parameter optimization unit that can adjust the text encoder learning rate (range 1e-6 to 1e-4) and the U-Net learning rate (range 1e-5 to 1e-3); a trigger word binding interface that associates and maps specific text tags with training data features. The LoRA model training module provides dataset annotation, trigger word binding, and training parameter optimization functions. Specifically, the LoRA model training module integrates a dataset annotation tool, a training parameter configuration interface, and a model testing environment, supports trigger word binding and dynamic adjustment of the learning rate, and is used to provide dataset annotation, trigger word binding, and training parameter optimization functions. The multimodal image processing module includes an intelligent face-swapping submodule that uses StyleGAN2 for facial feature encoding and transfer; a super-resolution reconstruction submodule that uses ESRGAN to achieve a 4x resolution increase; and a progressive inpainting algorithm that uses a multi-scale residual network to restore texture details in old photos. The multimodal image processing module implements local repainting based on an attention mechanism, intelligent image enlargement based on a generative adversarial network (GAN), and super-resolution inpainting based on a convolutional neural network (CNN), forming a multi-stage image processing pipeline that includes local repainting, intelligent image enlargement, super-resolution inpainting, and style transfer.

[0023] The content security review module includes: a text filtering layer that uses the BERT-based model to detect sensitive keywords such as those related to politics; an image review layer that uses YOLOv5 to identify illegal visual elements with a confidence threshold set to 0.85; and an audit tracking system that records the hash values and audit logs of all generated content. The content security review module uses natural language processing technology to filter input text for compliance and computer vision technology to identify the content of generated images. This module utilizes a dual-channel content security review mechanism, employing BERT text filtering and YOLOv5 image recognition technology.

[0024] This invention overcomes the technical defects of existing AI generation tools that cannot support user-defined models and dynamic expansion services. It integrates LoRA model training and online generation pipelines to support users' full-process closed-loop operations from data annotation to model deployment.

[0025] The implementation method of the Web-based intelligent image creation system of the present invention includes: S1. Build web application services Front-end: Vue.js 3 + Element Plus builds a responsive interface, supports multi-language input, and parameter visualization configuration (such as diffusion steps and classifier guidance ratio) Backend: Spring Boot framework implements RESTful API and integrates MyBatis ORM to operate MySQL 8.0 database Security module: JWT token authentication + OAuth 2.0 authorization protocol; S2. Use Kubernetes to build a Docker container cluster Docker container cluster: Schedule the text-to-image generation model (Stable Diffusion v2.1), image conversion model (ControlNet), and LoRA model through Kubernetes; Dynamic expansion mechanism: automatically adds container nodes based on GPU memory usage (threshold > 80%); S3. Build a distributed file storage system and a file server.

[0026] Distributed File System (DFS): Storage system model, user model (supports up to 10GB / file), LoRA model, training dataset, and generated works; File server: Nginx provides HTTP / HTTPS file access services, supporting breakpoint resuming and parallel downloading.

[0027] Specifically, if Figure 2 As shown, the construction (implementation) process of the Web-based intelligent image creation system of the present invention includes: 101. Build a web application service. The web application service adopts a front-end and back-end separation model. The front-end uses Vue3.js and Element Plus to build the front-end framework UI components, providing users with a smooth and intuitive responsive operation interface, facilitating data input, result display, and query.

[0028] The backend uses the Spring Boot framework to build its basic architecture and MyBatis as the persistence layer framework to access the database. The backend is responsible for receiving front-end interface data, processing business logic, connecting to and reading from the database, performing data calculation and storage, and returning processed data to the front-end page.

[0029] 102. Build a database. The database uses MySQL 8.0 to store system-related information, including user login information, user work creation information, text and image model information, LoRA model information, training services, and image creation service information.

[0030] 103. Use Kubernetes to build a docker container cluster. Run model images such as Wensheng graph model images, graph graph model images, super-resolution repair models, and LoRA models on the docker node. Docker nodes can dynamically expand node capacity based on platform concurrency and task types, and support dynamic allocation of computing resources based on load. After the model image is started, a node heartbeat message is sent to the database every 10 seconds. After the new node sends a heartbeat message for the first time, the database adds and saves the heartbeat message. After sending a heartbeat message, the database updates the heartbeat sending time of the node. The heartbeat information includes the node address, service type (Wensheng graph (such as Figure 1 In the text-generated graph model 1, text-generated graph model 2), graph-generated graph (as shown in the graph-generated graph model 1, graph-generated graph model 2), super-resolution, restoration, LoRA training (as shown in Figure 1 Lora training 1, Lora training 2), etc.), model name, sending time, etc., provide information query and address routing for the web backend to call related API services such as Wensheng graph, image processing and LoRA training.

[0031] 104. Build a distributed file storage system (DFS). This system is used to store various text-based graph models, graph-based graph models, LoRa models, and user-generated image artwork. For example, this system includes default models, user-uploaded models, user-uploaded datasets, and trained LoRa models. It supports multiple copies for redundant backup.

[0032] 105. Build a file server. Use nginx to build a file server and mount the corresponding distributed storage directory. Provide online file access, facilitating online access to image works and uploading and downloading model files.

[0033] The innovative features of the web-based intelligent image creation system of the present invention include: Web-based application-based text and image functions Front-end architecture: Vue.js 3 combined with Element Plus is used to build a responsive user interface, providing users with an intuitive and smooth operating experience.

[0034] Back-end architecture: Use the Spring Boot framework with MyBatis for data access, and use the MySQL database to store system-related information and operation records.

[0035] Image Creation Service: Docker containerized deployment of text-based graph models, graph-based graph models, and LoRA models ensures the stability and scalability of the system's image creation service. Users can guide the image generation process by selecting different image creation types and related models, entering prompt words and reverse prompt words, and setting parameters such as seed, step (number of diffusion model iterations), CFGScale (classifier free bootstrapping scale), image size, and number of generations.

[0036] User-defined model upload and selection Provides users with environment and computing power support, allowing users to upload their own Wensheng image models and LoRA models. When creating, users can choose to use the system's existing image creation models or upload their own image creation models or LoRA models to achieve a more desired creative effect.

[0037] LoRA model training function For users dissatisfied with existing LoRA models, the system provides a comprehensive LoRA model training mechanism. Users can upload an image collection, select a labeling algorithm, set a labeling threshold, and trigger words to annotate the dataset. They then select a training base model and set training parameters (such as the number of training rounds and save frequency) to initiate the training process. Once training is complete, users can publish the newly trained LoRA model for subsequent image creation.

[0038] Advanced image editing features In addition to the initial image generation, the system also provides a wealth of secondary creation options, such as partial redrawing, intelligent image expansion, intelligent extraction, intelligent face replacement, and intelligent removal. In addition, for old photos, the system supports super-resolution reconstruction and restoration to enhance the quality of old photos.

[0039] The business function modules of the Web-based intelligent image creation system of the present invention are mainly divided into four main modules: work square module, online image generation module, image creation module, and LoRA training module.

[0040] The Works Plaza module is used to display excellent user-created image works and basic information about the works, including prompt words, reverse prompt words, seed, step (number of diffusion model iterations), CFG Scale (classifier free guidance scale), image size, creation time, etc., to facilitate other creators to restore the works and find inspiration; The online image generation module is used for document generation and image editing. Figure 3 As shown in the figure, the main process of creating a cultural map includes: 201. Select a cultural image model: Select a basic model from the model list (natural scenery, bird photography, ancient Chinese clothing, Japanese anime, etc.).

[0041] 202. Select LoRA model: Select the style (LoRA model) of the image you want to create. The default is random.

[0042] 203. Enter the text description of the image you want to create. Text descriptions can be entered in multiple languages, and the backend will automatically recognize and translate it into English to call the text-to-image service. If you are not good at text descriptions or want a richer text description, turn on the prompt word optimization function. The backend will use the language model to automatically expand the text description to make it more rich and detailed.

[0043] 204. Set more parameters to guide the model to create images: Open the settings professional mode to set more parameters, such as image width and height, reverse prompt word, seed, step (diffusion model iteration steps), CFG Scale (classifier free guidance scale) and other parameters.

[0044] 205. Start creating raw images: Click the "Generate Now" button to begin creating images. The frontend initiates an HTTP request with the relevant parameters to the backend. The backend accepts and analyzes the parameters, and uses the model ID in the parameters to query the database for the node address of the model service. The backend then initiates an image creation request with the relevant parameters to the model creation service via WebClient (a new responsive, non-blocking HTTP client introduced in Spring Framework 5). After generating the image, the model creation service returns the image's Base64 format to the backend. The backend saves the image's Base64 format to the corresponding directory on the file server and generates an image URL to return to the frontend. The frontend page displays the image using the image URL. Excellent works can be published to the system with one click. After the administrator approves them, they will be automatically displayed in the Works Forum. Users can also download and save them locally.

[0045] Regarding content security, the following measures were implemented: Text content security audits were conducted on the input side, using the BERT model to detect whether Chinese text descriptions contain content that violates red lines, such as political or prohibited content. On the output side, security audits were conducted on AIGC-generated images, using YOLOv5 to identify illegal visual elements and detect whether AIGC-generated images contain content that violates regulations or is inappropriate for dissemination. This dual approach ensures that AIGC-generated content is secure, reliable, positive, and consistent with correct values.

[0046] In addition, if users have a good Wenshengtu model or LoRA model that they have used before, they can directly upload the Wenshengtu model or LoRA model through the system publishing function. After the administrator reviews and approves it, it can be directly run in the system and open to all platform users.

[0047] 206. Image Re-creation (Partial Redrawing, Image Expansion, Elimination, Face Swapping, etc.): Generated images can be further edited. Partial Redrawing: Select a portion of the image, enter a text description of the content you wish to redraw, and the backend will call the API provided by the redrawing model to regenerate the redrawn image and return it to the frontend for display. Intelligent Image Expansion: Set the width and height of the image area to be expanded, enter a text description of the content you wish to redraw, and the backend will call the API provided by the image expansion model to generate the expanded image and return it to the frontend for display. Intelligent Extraction: Select the image to be extracted, initiate an extraction request, and the backend will call the API provided by the extraction model to extract the image subject, set the background to transparent, and return it to the frontend. Intelligent Elimination: Select an image, enter a text description of the content to be eliminated, and the backend will call the API provided by the elimination model to generate the eliminated image and return it to the frontend for display. Intelligent Face Swapping: Use the StyleGAN encoder to transfer facial features. Select the image to be swapped, upload a new profile picture, and the backend will call the API provided by the face swapping model to generate the swapped image and return it to the frontend for display.

[0048] The image creation module is used for image stylization conversion and super-resolution restoration of old photos. For image stylization conversion, users upload the image to be converted and select the desired style, such as watercolor, cartoon illustration, 3D, Japanese anime, cyberpunk, or sketch. The backend then calls the API provided by the stylization conversion model to generate the converted image and return it to the frontend for display. For old photo restoration, super-resolution uses the ESRGAN network to increase the resolution by 4x. The restoration utilizes a progressive restoration algorithm and a multi-scale attention mechanism to restore details in old photos. Users upload the photo to be super-resolution restored, and the backend calls the API provided by the super-resolution restoration model to generate the restored image and return it to the frontend for display.

[0049] The LoRA model training module is used to provide LoRA model training functions. For some professional users, the existing LoRA models in the system may not meet their creative requirements. The system provides a complete and simple LoRA training service for some professional users. Users do not need to build a training environment or purchase a training server, and can easily train a series of LoRA models that meet their requirements. Figure 4 As shown, the specific steps of LoRA training are: 301. Create a new dataset: The user creates a new dataset and sets the dataset name.

[0050] 302. Upload the dataset that needs to be trained.

[0051] 303. Set the marking related parameters and mark the picture: select the marking algorithm, marking threshold, trigger words and other parameters for marking.

[0052] 304. Select the training base model and set the training parameters, such as the number of training rounds, the frequency of saving the LoRA model, text_encoder_lr and unet_lr (learning rate of the text encoder and U-Net model), and other related parameters.

[0053] 305. Start LoRA training. The training progress can be viewed in the task list. After the training is completed, you can view the list of trained LoRA models.

[0054] 306. The trained LoRA model can be published to the system for raw image testing. Users can directly test it with the base model in the online raw image module to see the raw image effect of the trained LoRA model. If the effect is not good, you can re-upload the dataset and retrain.

[0055] The present invention provides an AI image creation system that is extremely easy to use. Users do not need to install any software and can start creating by simply accessing it through a browser. The system adopts an intuitive and friendly interface design, supports simple parameter settings and automatic prompt word optimization, and even non-professional users can easily generate ideal pictures. It has built-in various practical functions such as partial redrawing, intelligent image expansion, face changing, etc., allowing users to further edit their works. It also provides old photo repair functions to reproduce precious memories. The content security mechanism ensures that the generated content is positive and healthy. In particular, users can upload custom models or train exclusive LoRA models, which greatly enhances the freedom and personalization of creation. The method for building a Web-based intelligent image creation system architecture of the present invention builds a complete ecosystem, and the method includes building a Web application service, building a docker container cluster, building a distributed file storage system, and building a file server. This series of measures not only ensures the stable operation of the system, but also provides users with an efficient, convenient and powerful image creation platform.

[0056] This invention achieves a 40% increase in service response speed through containerized deployment and supports thousands of concurrent tasks; the user-defined model function increases creative freedom by 300% and supports personalized model sharing; a one-stop multi-level image processing pipeline, including local redrawing, intelligent image expansion, super-resolution restoration and style transfer functions; the built-in LoRA training module reduces professional user training costs by 80%, shortens the training cycle to 30 minutes, and supports users' full-process closed-loop operations from data annotation to model deployment; the dual-channel content review mechanism of text input and image output achieves a 99.8% illegal content interception rate.

[0057] In summary, this invention not only represents a significant advancement in image editing technology, but also provides a new world of unlimited possibilities for creators. As technology continues to evolve and improve, this system will have broad application prospects in future digital art creation.

[0058] The above description is the best embodiment based on the concept and working principle of the invention. The above embodiment should not be understood as limiting the scope of protection of the present claims. Other embodiments and combinations of implementations of the present invention are also within the scope of protection of the present invention.

Claims

1. A Web-based intelligent image creation system, characterized in that: include: The web application service module uses the Vue.js 3 framework to build the front-end interactive interface, the Spring Boot framework to implement the back-end business logic, and the RESTful API for front-end and back-end communication; Model serving cluster, including Docker containers managed by Kubernetes to deploy text-to-image generation models, image-to-image conversion models, and low-rank adaptive models LoRA, supporting dynamic allocation of computing resources based on load; Distributed File Storage System (DFS), used to store user-defined models, training datasets, and generated works, supporting multiple copies of redundant backups; The LoRA model training module integrates dataset annotation tools, training parameter configuration interface, and model testing environment, and supports trigger word binding and dynamic adjustment of learning rate; Multimodal image processing module, including a local redrawing submodule based on an attention mechanism, an intelligent image expansion submodule based on a generative adversarial network (GAN), and a super-resolution restoration submodule based on a convolutional neural network (CNN); as well as The dual-channel content security review module uses natural language processing technology to filter the input text for compliance and uses computer vision technology to perform content recognition on generated images.

2. The Web-based intelligent image creation system according to claim 1, characterized in that: in, The model service cluster also includes: Heartbeat monitoring mechanism, where each container node sends status information including service type, model version, and resource utilization to the database every 10 seconds; and The service routing module dynamically generates a list of available service nodes based on heartbeat information and allocates generation tasks through a weighted round-robin algorithm.

3. The web-based intelligent image creation system according to claim 1, characterized in that: The LoRA model training module also includes: Automatic annotation unit, used to semantically annotate training images; Parameter optimization unit, used to adjust the text encoder learning rate and U-Net learning rate; and The trigger word binding interface is used to associate and map specific text tags with training data features.

4. The web-based intelligent image creation system according to claim 1, characterized in that: The multimodal image processing module further includes: The intelligent face-changing submodule uses StyleGAN2 for facial feature encoding and migration; A super-resolution reconstruction submodule that uses ESRGAN to achieve a 4x resolution improvement; and A progressive restoration algorithm that restores texture details in old photos through a multi-scale residual network.

5. The Web-based intelligent image creation system according to claim 1, characterized in that: The content security audit module includes: The text filtering layer uses the BERT-base model to detect politically sensitive keywords; Image review layer, using YOLOv5 to identify illegal visual elements; Audit trail system that records hash values and audit logs of all generated content.

6. The Web-based intelligent image creation system according to claim 1, characterized in that: The integrated LoRA model training module is used to provide data set annotation, trigger word binding and training parameter optimization functions.

7. The web-based intelligent image creation system according to claim 1, characterized in that: The dual-channel content security review module adopts a dual-channel content security review mechanism, using BERT text filtering and YOLOv5 image recognition technology.

8. The method for implementing the Web-based intelligent image creation system according to any one of claims 1 to 7, characterized in that: include: S1. Build web application services; S2. Use Kubernetes to build the Docker container cluster; as well as S3. Build the distributed file storage system and the file server, wherein the distributed file storage system is used to store various types of text-based graph models, graph-based graph models, LoRA models, and user-created image works, and use nginx to build the file server and mount the corresponding distributed storage directory.

Images