An AIGC-based interactive large-screen real-time drawing method and system

Abstract

The application discloses an interactive large-screen real-time drawing method and system based on AIGC, and mainly relates to the technical field of artificial intelligence generated content. The method comprises the following steps: establishing a real-time data connection between a mobile terminal and an interactive large screen; collecting drawing input data of a user through the real-time data connection, and performing structured preprocessing on the drawing input data to generate structured data; generating personalized prompt words based on a selected style of the user and historical behavior data of the user, and generating a stylized image through a diffusion model in combination with the generated structured data; dynamically assigning a stylized image generation task to a computing node; and pushing the generated stylized image to the interactive large screen in real time for display. The application has the beneficial effects that real-time synchronization of mobile terminal drawing and large screen display and personalized style generation are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence generated content, in particular to an interactive large screen real-time drawing method and system based on AIGC. BACKGROUND

[0002] With the rapid development of AIGC (Artificial Intelligence Generated Content) technology, its application in the field of image generation and artistic creation is increasingly widespread. The emergence of diffusion models such as Text-to-Image provides strong support for content creation. However, when applying AIGC technology to real-time interactive scenarios, especially real-time drawing involving cross-terminal collaboration, existing technologies still face significant challenges in real-time performance, personalization, and resource scheduling.

[0003] For example, existing intelligent copywriting or content recommendation systems (such as the contents disclosed in patents CN119783645B and CN120196815B) focus their core optimization on personalized generation or intent understanding based on historical data. Although these solutions have improved in terms of personalization, they typically handle offline or non-real-time tasks, and their model iteration and response patterns are difficult to meet the millisecond-level low-latency feedback required for real-time drawing. On the other hand, some technical solutions (such as the content disclosed in patent CN119477658A) focus on resource scheduling issues in real-time inference, allocating computing resources through a scheduling platform to avoid congestion. However, such methods focus on resource allocation itself and fail to deeply integrate with the AIGC generation process, especially lacking consideration of real-time preprocessing (such as skeleton extraction) of user drawing input to optimize the load of the generation model. Moreover, when dealing with high-concurrency requests, it is difficult to ensure both low latency and high generation quality. In addition, current AIGC text-to-image technology generally has limited controllability of generated content and high computational resource consumption.

[0004] Therefore, there is an urgent need for an interactive large screen real-time drawing method and system based on AIGC, which can effectively solve the technical problem of achieving low-latency, high-personalization experience in the full link from mobile input to AIGC real-time generation and display in an interactive large screen scenario. SUMMARY

[0005] The present application aims to provide an interactive large screen real-time drawing method and system based on AIGC, which realizes real-time synchronization of mobile drawing and large screen display and personalized style generation.

[0006] To achieve the above-mentioned purpose, the present application realizes the following technical solutions:

[0007] On the one hand, the present application provides an interactive large screen real-time drawing method based on AIGC, comprising the following steps:

[0008] Step S1: Establish a real-time data connection between the mobile terminal and the interactive large screen;

[0009] Step S2: Collect user drawing input data through the real-time data connection, and perform structured preprocessing on the drawing input data to generate structured data;

[0010] Step S3: Generate personalized prompt words based on the user-selected style and user historical behavior data, and generate stylized images through a diffusion model combined with the structured data generated in step S2;

[0011] Step S4: Dynamically assign the stylized image generation task in step S3 to a computing node;

[0012] Step S5: Real-time push the generated stylized images to the interactive large screen for display.

[0013] Preferably, step S1 includes:

[0014] Step S11: After the application on the interactive large screen is started, a unique session identifier is applied to the server, and a two-dimensional code containing the session identifier is generated and displayed;

[0015] Step S12: The user scans the two-dimensional code in step S11 through the mobile terminal, and after the drawing board application on the mobile terminal is started, the session identifier is checked with the server, and a point-to-point real-time data channel bound with the interactive large screen is successfully established;

[0016] The real-time data channel in step S12 uses the WebSocket protocol.

[0017] Preferably, in step S2, the structured preprocessing of the drawing input data includes skeleton key point extraction and line contour extraction.

[0018] The skeleton key point extraction includes:

[0019] Input continuous stroke coordinate sequence from mobile terminal , wherein is the coordinate;

[0020] Based on the timestamp and the coordinate distance threshold, the discrete coordinate points are aggregated into independent stroke objects ;

[0021] The Douglas-Peucker algorithm is used to simplify the trajectory of each stroke and retain the core shape;

[0022] The simplified stroke set is input into a lightweight graph convolution network to output a fixed-dimensional skeleton key point feature vector ;

[0023] The line profile extraction includes:

[0024] The original stroke coordinates are rendered to a low-resolution grayscale bitmap .

[0025] The Canny edge detector is applied for edge enhancement, and a line binary graph is output .

[0026] Preferably, the graph convolution network in the skeleton key point extraction is pre-trained by human or animal simple sketch data, learns the mapping from simple lines to standard skeleton structure, and compresses the model to improve the inference speed through model distillation and quantization technology.

[0027] Preferably, the Canny edge detector in the line profile extraction specifically performs the following steps:

[0028] Smooth using a Gaussian filter: .

[0029] Calculate the gradient amplitude and direction , wherein and are the convolution results of the Sobel operator;

[0030] Adopt non-maximum suppression to refine the edges;

[0031] Adopt double-threshold detection and edge connection to finally output a line binary graph .

[0032] Preferably, step S3 includes:

[0033] The personalized prompt word generator generates style modification words according to the user-selected basic style and user historical behavior feature data, and generates the final personalized style prompt word by combining the style modification words .

[0034] The style conversion model is realized based on the Control Net diffusion model, and the structured data obtained by preprocessing is taken as the control condition and input into the diffusion model together with the personalized style prompt word;

[0035] The diffusion model adopts Stable Diffusion as the basic model, and the Control Net branch takes the conditional graph as the input, extracts the conditional features through the encoder layer, and adds the feature map output by the Control Net branch to the feature map of the main branch of the UNet at each down-sampling block of the UNet.

[0036] The denoising process removes noise Start, through Sub-iteration denoising, each step The predicted noise Controlled by joint conditions:

[0037] ;

[0038] Where, The noisy latent variable of the first Step, Text prompt words Text embedding obtained by CLIP text encoder, Condition features extracted by Control Net for conditional graph .

[0039] Preferably, the intelligent computing power scheduling center in step S4 performs the following operations:

[0040] Maintain real-time updated capability and load profile for each computing node Calculate the comprehensive scheduling priority score , Where The node computing power coefficient, Real-time comprehensive load score;

[0041] The calculation formula is:

[0042] ;

[0043] Where , , CPU, GPU, memory usage, The number of queued tasks on the node, Network delay, The current measured computing power utilization rate of the node, , , , , , Adjustable weight coefficient;

[0044] When a new task arrives, the scheduling center pre-selects nodes with high load or unreachable network according to the task complexity classification, and selects nodes with the highest comprehensive scheduling priority score To execute the task:

[0045] ;

[0046] When the scores are similar, the edge node with a closer physical location is selected preferentially, and the service continuity is realized through task heartbeat and migration, and the elastic scaling mechanism.

[0047] Preferably, step S5 comprises:

[0048] After the style conversion engine generates the image, the processing node first down-samples the image into a low-resolution preview image and preferentially pushes it to the interactive large screen to realize millisecond-level instantaneous preview;

[0049] Then, the complete high-definition original image is transmitted in the background for seamless replacement;

[0050] The interactive large screen renders immediately after receiving the preview image and automatically switches when the high-definition image arrives.

[0051] On the other hand, the present application also provides an interactive large screen real-time drawing system based on AIGC, which is used to realize the interactive large screen real-time drawing method based on AIGC as described above, comprising:

[0052] A mobile terminal for providing a drawing interface and collecting user drawing input data;

[0053] An interactive large screen for displaying the drawing process, generating a result preview, and the final high-definition image;

[0054] A server processing cluster for session management, real-time data reception, structured preprocessing, and personalized prompt word generation;

[0055] An intelligent computing power scheduling center for monitoring the real-time state of the computing nodes and dynamically allocating style image generation tasks;

[0056] Edge computing nodes and cloud GPU computing clusters for executing preprocessing and diffusion model inference tasks;

[0057] User data and model library for storing user historical behavior data, personalized portraits, and model parameters.

[0058] Preferably, the server processing cluster comprises a preprocessing unit for executing skeleton key point extraction and line contour extraction;

[0059] The personalized prompt word generator dynamically generates enhanced prompt words using a pre-trained language model;

[0060] The style conversion model is realized based on the Control Net diffusion model, combining structured data and personalized prompt word generation to generate stylized images;

[0061] The intelligent computing power scheduling center performs task allocation by comprehensively scheduling priority score

[0062] ​The interactive large screen adopts a two-stage transmission optimization strategy, first previewing a low-resolution image, and then replacing it with a high-definition image.

[0063] Compared with the prior art, the present application has the following advantages:

[0064] 1. Strong real-time interaction: The present application realizes second-level synchronization between mobile end drawing and large screen display through real-time data channel and computing power scheduling, ensuring the smoothness of the creation process.

[0065] 2. High degree of personalization: The present application generates personalized prompt words based on user real-time drawing content and historical preferences, enabling AIGC to output precise matching of user artistic style.

[0066] 3. Good system stability: The present application adopts an intelligent computing power scheduling scheme to dynamically allocate computing resources, ensuring service stability and low latency in high-concurrency scenarios.

[0067] 4. Good user experience: The present application simplifies complex AIGC technology into intuitive drawing interaction, significantly reducing the use threshold and enhancing the joy of creation. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 is a method flowchart of an embodiment of the present application;

[0069] Figure 2 is a system structure schematic diagram of an embodiment of the present application;

[0070] Figure 3 is an image generation data flow diagram of an embodiment of the present application;

[0071] Figure 4 is an interactive large screen real-time display and feedback data flow diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0072] The present application will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present application and not to limit the scope of the present application. In addition, it should be understood that those skilled in the art can make various modifications or changes to the present application after reading the content taught by the present application, and these equivalent forms also fall within the scope defined by the present application.

[0073] In the present application, terms such as "up", "down", "left", "right", "front", "back", "vertical", "horizontal", "side", "bottom", etc. indicate the orientation or positional relationship shown in the drawings, which is only a relationship word determined for the convenience of describing the structural relationship of the components or elements of the present application, and cannot be understood as a limitation of the present application.

[0074] Embodiment:

[0075] As Figure 1 shown, the embodiment provides an AIGC-based interactive large-screen real-time drawing method, mainly including the following steps:

[0076] Step S1: Establish a real-time data connection between the mobile terminal and the interactive large screen;

[0077] Step S2: Collect user drawing input data through real-time data connection, and perform structured preprocessing on the drawing input data to generate structured data;

[0078] Step S3: Generate personalized prompt words based on the user-selected style and user historical behavior data, and generate stylized images through a diffusion model combined with the structured data generated in step S2;

[0079] Step S4: Dynamically assign the stylized image generation task in step S3 to a computing node;

[0080] Step S5: Real-time push the generated stylized images to the interactive large screen for display.

[0081] As Figure 2 shown, the embodiment also provides an AIGC-based interactive large-screen real-time drawing system, comprising:

[0082] Mobile terminal: In this embodiment, it is a user's smartphone, which provides a drawing interface through a WeChat mini-program. When the mini-program in the user's smartphone is started, it establishes a real-time data channel (such as WebSocket) with the server by scanning the large-screen QR code;

[0083] Interactive large screen: In this embodiment, it is a smart TV, which is responsible for displaying the drawing process, generating result preview and final high-definition image, and providing a user feedback interface;

[0084] Server processing cluster: In this embodiment, it is the core processing unit of the system, responsible for session management, real-time data reception, preprocessing (skeleton extraction, line contour generation), personalized prompt word generation, and calling AIGC engine for image generation;

[0085] Intelligent computing power scheduling center: In this embodiment, it is responsible for monitoring the real-time state of all computing nodes (including cloud GPU cluster and edge node), and according to the dynamic load evaluation result, the generation task is scheduled to the optimal node for execution;

[0086] Edge computing node: Distributed in the user's local network or near-field environment, it is a computing device (such as a home computer, a local server) that can undertake part or all of the preprocessing tasks and lightweight inference tasks to reduce the load and transmission delay of the core cloud

[0087] Cloud GPU computing cluster: composed of multiple high-performance GPU nodes, responsible for executing computationally intensive AIGC model inference tasks such as joint generation of Stable Diffusion and Control Net;

[0088] User data and model library: stores user historical behavior data, personalized portraits, exclusive style model parameters, and session records, supporting personalized generation and continuous optimization of the system.

[0089] In combination with the system composition in this embodiment, the AIGC-based interactive large-screen real-time drawing method in this embodiment is described in detail.

[0090] Step S1: cross-device binding and interaction start:

[0091] After the application on the interactive large screen (smart TV) is started, a unique session identifier is applied to the server, and a two-dimensional code containing the identifier is generated and displayed;

[0092] The user scans the two-dimensional code through the WeChat on the mobile terminal (user's smart phone), and after the mini-program drawing board application is started, the captured session identifier is verified with the server, and a point-to-point real-time data channel (preferably WebSocket protocol) bound to the specific large screen is successfully established;

[0093] The drawing board application provides a touch drawing interface and built-in multiple preset styles (such as "anime style", "oil painting", "cyberpunk") for user selection.

[0094] Step S2: real-time data collection and structured preprocessing:

[0095] As shown in Figure 3 , the process of converting raw strokes to structured data, user drawing operations on the drawing board (such as stroke trajectory, coordinates, pressure information, timestamp) are encapsulated into data packets in real time and uploaded to the server through the data channel;

[0096] After the server's preprocessing unit receives the raw stroke data, it performs key structured preprocessing operations, including:

[0097] Skeleton key point extraction: for drawings involving characters, animals, and other subjects, the following algorithm is used to estimate and output skeleton key point coordinate data from the stroke sequence:

[0098] (1) Input: continuous stroke coordinate sequence from the mobile terminal , where is the coordinate;

[0099] (2) Stroke aggregation: based on timestamp and coordinate distance threshold, discrete coordinate points are aggregated into independent stroke objects ;

[0100] (3) Graph abstraction: Douglas-Peucker algorithm is used to simplify the trajectory of each stroke, reduce redundant points, and retain the core shape. The algorithm finds the point farthest from a line segment by recursion. If the distance is greater than the threshold , the line segment is divided at the point, and finally all key turning points are retained;

[0101] (4) Skeleton inference: the simplified stroke set is input into a lightweight graph convolutional network (GCN). The network regards each stroke as a node in the graph and the spatial relationship between nodes as an edge. Through graph convolution operation, the neighborhood information is aggregated, and finally a fixed-dimensional skeleton key point feature vector is output . This GCN is pre-trained on a large amount of human / animal sketch data, learns the mapping from simple lines to standard skeleton structure, and compresses the model through model distillation and quantization techniques to improve the model inference speed;

[0102] Line contour extraction: the following algorithm is used to extract the main contour from the stroke and generate a clean line binary image:

[0103] (1) Generate bitmap: render the original stroke coordinates to a low-resolution (e.g. 64x64) grayscale bitmap ;

[0104] (2) Edge enhancement: apply Canny edge detector:

[0105] First, use a Gaussian filter for smoothing: ;

[0106] Calculate the gradient amplitude , direction , where and are the convolution results of the Sobel operator;

[0107] Use non-maximum suppression (NMS) to refine the edges;

[0108] Use double-threshold detection and edge connection to finally output the line binary image ;

[0109] This preprocessing process can be decided by the server and issued to edge devices with computing power (such as home computers, local servers) to execute according to system load, to share the computing pressure of the core server and reduce transmission delay.

[0110] Step S3: personalized prompt word generation and controllable style conversion

[0111] The personalized prompt generator dynamically generates enhanced prompts according to the user's selected "target style" and their historical behavior profile (such as past viewing records, art browsing tags);

[0112] The specific logic is as follows:

[0113] (1) Assuming the user selects the basic style as (such as "Anime style");

[0114] (2) The system extracts historical behavior feature data (such as content preferences, family roles, etc.) from the user's profile library;

[0115] (3) Use a pre-trained language model (such as CLIP text encoder) to extract user behavior features and generate more rich modifiers (such as: Disney, bright colors);

[0116] (4) Combine these style modifiers with the user's selected basic style name to form the final personalized style prompt, and the final prompt is: (such as: Anime style, Disney, bright colors);

[0117] The style conversion model is implemented based on the Control Net diffusion model. First, load the pre-trained text-to-image diffusion model, and use the preprocessed structured data (skeleton key point coordinate data or line binary graph) as the control condition, input the diffusion model together with the personalized style prompt, and guide the model to generate the specified style on the given composition. This ensures that the generated content not only meets the user's drawing intention, but also meets their personalized aesthetic preferences. The specific steps are as follows:

[0118] (1) Model architecture: Use StableDiffusion (SD) as the base model and inject a ControlNet branch. The branch takes the preprocessed condition graph (such as line graph ) as input and extracts condition features through a series of encoder layers (same structure as SD's UNet encoder but with zero initialization weights);

[0119] (2) Fusion process: At each downsampling block of the SD UNet, the feature map output by the Control Net branch is added element-wise to the feature map of the UNet main branch, thereby injecting line or skeleton structure information into the generation process;

[0120] (3) Denoising process: The generation follows the principles of the diffusion model, starting from noise , and going through iterations of denoising, with the predicted noise at each step controlled by the joint condition:

[0121] ;

[0122] wherein:

[0123] is the noisy latent variable of the step, is the text prompt word encoded by the CLIP text encoder, is the conditional feature extracted by the Control Net for the conditional graph;

[0124] (4) Loss function (for incremental learning): when fine-tuning the model, the LDM (Latent Diffusion Model) loss function is used:

[0125] ;

[0126] wherein: is the latent encoding of the real image, is the real noise, denotes expectation, by minimizing this loss, the model learns to generate images that better meet the user's preferences under given personalized prompts and structural constraints;

[0127] To achieve continuous personalization, the system sets up a model optimization mechanism: when the user confirms a generation result, the result and its corresponding input data will be stored in the user's personal data set. The system regularly uses this data set to incrementally fine-tune the user's exclusive style model, so that its output better meets the user's evolving preferences.

[0128] Step S4: Intelligent computing power scheduling and task execution:

[0129] The intelligent computing power scheduling center, as the core hub of the system, is responsible for monitoring, evaluating and dynamically allocating global computing resources. It maintains a heterogeneous resource pool containing cloud high-performance GPU inference nodes and edge computing nodes (such as home computers and local servers), and monitors the multi-dimensional performance indicators of each node in real time;

[0130] (1) Dynamic load evaluation and node capability profiling:

[0131] The scheduling center maintains a real-time updated capability and load profile for each computing node , and calculates a comprehensive scheduling priority score , which takes into account both the node's static computing power and dynamic load. The calculation formula is as follows:

[0132] ;

[0133] Wherein: is the node computing power coefficient, which is a normalized constant value determined according to the node hardware configuration (such as GPU model, video memory size, CPU core number), used to reflect the inherent computing power of the node, for example, the node equipped with H100 GPU can be set to 2.0, while the node equipped with RTX4090 is set to 1.5, and the node containing only CPU is set to 0.8;

[0134] is the real-time comprehensive load score, and its calculation formula is optimized as:

[0135] ;

[0136] Formula variable description:

[0137] , , : CPU, GPU, memory usage rate (between 0 and 1);

[0138] : the number of tasks currently queued for processing on the node;

[0139] : network delay from the scheduling center to the node (unit: ms), normalized;

[0140] : the current measured computing power utilization rate of the node, which is dynamically obtained by running a standard benchmark test program, reflecting the actual effective computing power of the node under the current load;

[0141] , , , , , are adjustable weight coefficients, which support adaptive adjustment according to the overall load status of the system and the type of task (for example, in high concurrency period, increase the weight of and , pay more attention to queuing and actual computing power);

[0142] (2) Adaptive scheduling decision:

[0143] When a new style generation task arrives, the scheduling center executes the following decision-making process:

[0144] Task classification: according to the task complexity (such as whether it needs to extract bones, generate image resolution), it is divided into high, medium and low priority;

[0145] Node pre-screening: filter out nodes with too high load (such as ) or a node that is unreachable on the network;

[0146] Optimal node selection: among the pre-screened node pool, the node with the highest comprehensive scheduling priority score is selected to execute the task, which achieves the goal of "assigning the most suitable task to the most idle and strongest node"; The highest node to execute the task, this strategy achieves the goal of "assigning the most suitable task to the most idle and strongest node";

[0147] ;

[0148] Priority and load balancing: in the case of similar scores, preferentially select the edge node that is closer in physical location to the task initiator user to reduce network delay, while effectively avoiding the accumulation of requests to a single node by introducing and dynamic weights, achieving efficient load balancing;

[0149] (3) Task fault tolerance and elastic scaling:

[0150] Task heartbeat and migration: the scheduling center and the computing nodes maintain a heartbeat mechanism, if a node failure or task timeout is detected, the task is automatically rescheduled to other healthy nodes to ensure service continuity;

[0151] Elastic scaling: when the average load of all nodes in the resource pool continuously exceeds the threshold, automatically trigger the horizontal expansion mechanism to apply for new computing nodes from the cloud GPU cluster to join the resource pool; when the load decreases, automatically release idle resources to save costs.

[0152] Step S5: Real-time result pushing and immersive display:

[0153] As Figure 4 shown in the figure, the interactive large screen displays and feeds back the data flow diagram in real time, and the style conversion engine generates images, and the processing node immediately executes the two-stage transmission optimization strategy:

[0154] (1) First, the image is quickly down-sampled into a low-resolution preview image and is preferentially pushed to the interactive large screen, achieving millisecond-level instantaneous preview;

[0155] (2) Then, continue to transmit the complete high-definition original image in the background for seamless replacement, the large screen display module receives the preview image and immediately renders, and automatically switches when the high-definition image arrives;

[0156] The whole process, from user drawing to final high-definition image display, the delay is controlled within seconds (preferably <3 seconds), through this "preview first, high-definition later" mechanism, the high synchronization of drawing process and visual effect is ensured, achieving the immersive creation experience of "pen following thought, screen following pen drawing".

[0157] Through the system architecture, data flow and interaction logic shown in the above embodiments and in combination with the various drawings, the application can realize that the whole process delay control from the user starting to draw to the high-definition image display on the large screen is within 3 seconds, the system stability is ensured through intelligent scheduling in a high-concurrency scene, and the user satisfaction is continuously improved through personalized prompts and model fine-tuning.

[0158] The above is a specific description of the preferred implementation of the application, but the application is not limited to the described embodiments. Those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the application. These equivalent modifications or replacements are all included in the scope defined by the claims of the present application.

Claims

1. A real-time interactive large-screen drawing method based on AIGC, characterized in that, Includes the following steps: Step S1: Establish a real-time data connection between the mobile terminal and the interactive large screen; Step S2: Collect the user's drawing input data through real-time data connection, and perform structured preprocessing on the drawing input data to generate structured data; Step S3: Generate personalized prompts based on the user's selected style and historical user behavior data, and combine them with the structured data generated in step S2 to generate a stylized image through a diffusion model; Step S4: Dynamically assign the stylized image generation task from step S3 to the computing nodes; Step S5: Push the generated stylized image to the interactive large screen for display in real time; Step S3 includes: Personalized prompt generator based on the user's selected base style Based on user historical behavior data, style modifiers are generated using a pre-trained language model and combined to form the final personalized style prompts. ; The style transfer model is based on the Control Net diffusion model, using preprocessed structured data as control conditions, which are input into the diffusion model along with personalized style prompts. The diffusion model uses StableDiffusion as the base model. The Control Net branch takes the conditional map as input, extracts conditional features through the encoder layer, and adds the feature map output by the Control Net branch to the feature map of the main branch of UNet element by element at each downsampling block of UNet. The noise reduction process starts from the noise Beginning, then Each iteration of noise reduction, each step Predicted noise Controlled by joint conditions: ； in, It is the first The step with noisy latent variables, It is a text prompt word Text embedding obtained after CLIP text encoder It is a Control Net conditional graph Extracted conditional features; The intelligent computing power scheduling center performs the following operations in step S4: For each computing node The system maintains real-time updated capabilities and load profiles, and calculates comprehensive scheduling priority scores. ,in This represents the node computing power coefficient. For real-time comprehensive load score; The calculation formula is: ； in , , These are CPU, GPU, and memory usage, respectively. This represents the number of tasks queued on the node. Due to network latency, This represents the current measured computing power utilization rate of the node. , , , , , These are adjustable weighting coefficients; When a new task arrives, the scheduling center categorizes it based on task complexity, pre-screens nodes with excessive load or unreachable networks, and selects nodes based on their comprehensive scheduling priority score. The highest node executes the task: ； When scores are similar, edge nodes with closer physical locations are prioritized, and service continuity is achieved through task heartbeat and migration, as well as elastic scaling mechanisms.

2. The interactive large-screen real-time drawing method based on AIGC according to claim 1, characterized in that, Step S1 includes: Step S11: After the application on the interactive screen starts, it requests a unique session identifier from the server and generates a QR code containing the session identifier for display. Step S12: The user scans the QR code in step S11 through a mobile terminal. After the drawing board application on the mobile terminal is launched, the session identifier is verified with the server, and a point-to-point real-time data channel bound to the interactive screen is successfully established. The real-time data channel in step S12 uses the WebSocket protocol.

3. The interactive large-screen real-time drawing method based on AIGC according to claim 1, characterized in that, In step S2, the structured preprocessing of the drawing input data includes: extraction of skeletal key points and extraction of line contours; The extraction of skeletal key points includes: Input a continuous stroke coordinate sequence from the mobile device ,in Coordinates; Based on timestamps and coordinate distance thresholds, discrete coordinate points are aggregated into independent stroke objects. ; The Douglas-Peucker algorithm is used to simplify the trajectory of each stroke while preserving the core shape; The simplified stroke set is input into a lightweight graph convolutional network, which outputs fixed-dimensional skeletal keypoint feature vectors. ; The line contour extraction includes: Original stroke coordinates Render to low-resolution grayscale bitmap superior; Edge enhancement is performed using the Canny edge detector, and a binary line image is output. .

4. The interactive large-screen real-time drawing method based on AIGC according to claim 3, characterized in that, The graph convolutional network used in the skeletal key point extraction is pre-trained with human or animal sketch data to learn the mapping from simple lines to standard skeletal structures, and the model is compressed through model distillation and quantization techniques to improve inference speed.

5. The interactive large-screen real-time drawing method based on AIGC according to claim 3, characterized in that, The Canny edge detector in the line contour extraction process specifically performs the following steps: Smoothing using a Gaussian filter: ; Calculate gradient magnitude and direction , ,in and It is the convolution result of the Sobel operator; Non-maximum suppression is used to refine the edges; The method employs dual threshold detection and edge connection to ultimately output a binary line image. .

6. The interactive large-screen real-time drawing method based on AIGC according to claim 1, characterized in that, Step S5 includes: After the style transfer engine generates an image, the processing node first downsamples the image into a low-resolution preview image and pushes it to the interactive large screen first, achieving millisecond-level instant preview; Then, the complete high-definition original image is transmitted in the background for seamless replacement; The interactive large screen renders the preview image immediately upon receiving it, and automatically switches to a high-resolution image upon arrival.

7. An interactive large-screen real-time painting system based on AIGC, used to implement the interactive large-screen real-time painting method based on AIGC as described in any one of claims 1-6, characterized in that, include: Mobile terminals are used to provide a drawing interface and collect users' drawing input data. An interactive large screen is used to display the painting process, preview the generated results, and the final high-definition image; The server-side processing cluster is used for session management, real-time data reception, structured preprocessing, and personalized prompt word generation. The intelligent computing power scheduling center is used to monitor the real-time status of computing nodes and dynamically allocate stylized image generation tasks. Edge computing nodes and cloud GPU computing clusters are used to perform preprocessing and diffusion model inference tasks; User data and model library, used to store user historical behavior data, personalized profiles and model parameters.

8. The interactive large-screen real-time drawing system based on AIGC according to claim 7, characterized in that, The server-side processing cluster includes a preprocessing unit for performing skeletal keypoint extraction and line contour extraction; The personalized prompt word generator dynamically generates enhanced prompt words using a pre-trained language model; The style transfer model is based on the Control Net diffusion model and combines structured data and personalized prompts to generate stylized images. The intelligent computing power scheduling center uses a comprehensive scheduling priority score. Assign tasks; The interactive large screen adopts a two-stage transmission optimization strategy, first previewing a low-resolution image and then replacing it with a high-resolution image.

Images