A hybrid content generation scheduling method and system for resource-constrained devices

Abstract

The application discloses a hybrid content generation scheduling method and system for resource-limited devices, and relates to the technical field of artificial intelligence optimization. The scheduling method comprises the following steps: receiving a user content generation request; evaluating the current resource state of the system; dynamically selecting a preset resource library or starting an on-demand generation engine to obtain content through a strategy scheduling algorithm; rendering and adapting the output content; and outputting to a terminal display module. The application also provides a hybrid content generation scheduling system for implementing the above method. The application solves the technical problem that consumer-level devices are difficult to balance the real-time performance, high quality and personalized requirements of content generation through intelligent scheduling, and is especially suitable for embedded intelligent interactive devices with limited computing power and memory.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence optimization technology, specifically to a hybrid content generation and scheduling method and system for resource-constrained devices. Background Technology

[0002] The development of generative AI technology has provided powerful capabilities for dynamically creating digital content. Meanwhile, emerging interactive technologies such as light field displays place higher demands on the quality and real-time performance of content generation. However, deploying generative models on consumer-grade embedded devices (such as intelligent robots and augmented reality glasses) faces significant challenges: real-time generation of high-quality content requires extremely high computing power, leading to high device costs, high power consumption, and significant response latency, severely impacting user experience; furthermore, content relying entirely on random generation lacks consistency and reliability, making it difficult to meet the accuracy requirements of specific scenarios (such as precise molecular model display in educational tutoring).

[0003] Existing technologies offer partial solutions. For example, patent document CN119693580A, a light field data acquisition and processing method based on artificial intelligence, calculates surface normal vectors by processing and transforming illumination images. Patent document CN115953780A, a method for constructing multi-dimensional light field complex scene graphs based on multi-view information fusion, generates multi-dimensional semantic scene graphs. However, these existing technologies either focus on low-level data acquisition and processing or on the modeling and generation of complex scenes, and neither addresses the core issue of how to intelligently select the optimal generation path for different types of content requests in a resource-constrained environment.

[0004] Current solutions mostly adopt a single mode: either relying entirely on pre-built content libraries, which guarantees real-time performance but lacks personalization and flexibility; or relying entirely on real-time cloud generation, which offers high flexibility but is affected by network latency and stability, and makes it difficult to protect user privacy. Therefore, there is an urgent need for an intelligent content generation scheduling method that can dynamically select the optimal content generation path based on task requirements, context, and system resource status, providing highly personalized and accurate content output while ensuring real-time performance, stability, and low power consumption. Summary of the Invention

[0005] The purpose of this invention is to provide a hybrid content generation scheduling method and system for resource-constrained devices. Through intelligent scheduling, it solves the technical problem that consumer-grade devices are unable to simultaneously meet the requirements of real-time, high-quality and personalized content generation, and is especially suitable for embedded intelligent interactive devices with limited computing power and memory.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0007] In a first aspect, the present invention provides a hybrid content generation and scheduling method for resource-constrained devices, comprising the following steps:

[0008] S1. Receive user content generation requests, parse the requests, and obtain the request type, content topic, and complexity score; the request type is simple query, complex creation, or precise generation.

[0009] S2. Based on the parsing results, evaluate the current resource status of the system, call the strategy scheduling algorithm, and dynamically select the path between the preset resource library and the on-demand generation engine.

[0010] S3. If the strategy scheduling algorithm decides to call the preset resource library, then perform multi-level matching and retrieval based on the content topic and output the optimal preset content.

[0011] If the strategy scheduling algorithm decides to activate the on-demand generation engine, it will allocate computing resources and call the generation model to generate the target content based on the content topic and complexity score.

[0012] S4. Render and optimize the optimal preset content or target content and adapt its format, then output it to the terminal display module.

[0013] Furthermore, in step S1, when parsing the request, a natural language processing model is used to parse the request and extract the key instruction entities in the request;

[0014] Determine the request type based on a predefined rule base;

[0015] Based on the abstraction level of the instruction entity and the required computing resources, a scoring model is used to quantify the complexity score.

[0016] Furthermore, the expression for the scoring model is:

[0017] ;

[0018] Where C is the complexity score, L is the abstraction level quantification value of the instruction entity, R is the estimated computational resources required, and w1 and w2 are preset weight coefficients.

[0019] Furthermore, in step S2, the decision logic of the strategy scheduling algorithm is based on a multi-dimensional evaluation matrix. The input variables of the multi-dimensional evaluation matrix include: the complexity score of the request; the matching confidence of the request content in the preset resource library; the current available computing resources of the system (CPU / GPU utilization, memory availability); network bandwidth and latency status (used to evaluate the feasibility of cloud generation); the estimated generation time and power consumption of the on-demand generation engine; and the output variable is a decision path that can be chosen from two options.

[0020] Furthermore, the policy scheduling algorithm adopts a threshold-based decision tree model, which is implemented as follows:

[0021] If the request type is determined to be "exactly generated" or the matching confidence is higher than the first set threshold (Th_predefined), then the default resource library will be called first.

[0022] If the request type is determined to be "complex creation" or the matching confidence is lower than the second set threshold (Th_generate), it is recommended to call the preset resource library;

[0023] If the request type is determined to be complex creation or the matching confidence is higher than the second set threshold (Th_generate), and the system's current available resources are higher than the resource security threshold (Th_resource), then the on-demand generation engine will be activated first.

[0024] If the system's current available resources are below the resource safety threshold (Th_resource), a forced decision will be made to call the preset resource library or return a resource shortage waiting instruction.

[0025] Furthermore, in step S3, the generated model is a lightweight generated model or a cloud service.

[0026] Furthermore, in step S3, multi-level matching and retrieval include:

[0027] First-level matching: Matches content themes (such as "chemistry", "biology", "art style") to first-level category nodes;

[0028] Second-level matching: Calculate vector similarity under the second-level category node based on the requested attribute tags (such as "water molecule" or "cartoon");

[0029] Search optimization: Based on the collaborative filtering algorithm, it recommends preset content with high relevance according to historical interaction data.

[0030] Furthermore, in step S3, the on-demand generation engine has a modular structure, including:

[0031] The image generation unit generates lightweight 3D virtual images based on text descriptions.

[0032] The formula parsing unit converts abstract information such as mathematical formulas and chemical molecular formulas into three-dimensional geometric structures that can be rendered.

[0033] The style transfer unit transforms the artistic style of existing preset content;

[0034] The generation process adopts a "cloud-edge" collaborative computing framework. Lightweight models run on local devices, while large models call cloud APIs and perform localized post-processing on the generated results.

[0035] Furthermore, in step S4, rendering optimization and format adaptation include:

[0036] The model is lightweighted, reducing the number of faces and texture resolution to a level that allows for smooth rendering on the target screen;

[0037] Material and lighting effects are remapped to adapt the rendering pipeline to specific display hardware;

[0038] The orientation and scale of the content are automatically adjusted based on the optimal viewing angle of the display device.

[0039] Secondly, the present invention also provides a hybrid content generation scheduling system for resource-constrained devices, for implementing the above-described hybrid content generation scheduling method for resource-constrained devices, the hybrid content generation scheduling system comprising:

[0040] The request parsing module is used to receive and parse the user's content generation request, and output the request type, content topic, and complexity score;

[0041] The strategy scheduling module, which has an embedded strategy scheduling algorithm, is used to make content generation path decisions based on the request parsing module.

[0042] The preset resource library module is used to store, index, and manage preset 3D models, animations, and special effects content, and to perform retrieval tasks;

[0043] The on-demand generation engine module is used to generate target content by calling local or cloud models after receiving instructions;

[0044] The rendering adaptation module is used to optimize the generated or retrieved content to adapt to the output requirements of the terminal display module.

[0045] The terminal display module is used to display the generated content.

[0046] Furthermore, the hybrid content generation system also includes a learning optimization module, which records each content request and its corresponding path decision and result feedback; based on the recorded data, a reinforcement learning algorithm is used to iteratively optimize the decision threshold and weight parameters in the strategy scheduling algorithm.

[0047] The present invention has the following beneficial effects:

[0048] 1. Intelligent dynamic scheduling for efficient resource utilization: Through innovative strategy scheduling algorithms, the optimal content generation path is dynamically selected based on the complexity of real-time requests and the status of system resources, effectively solving the contradiction between "high quality" and "real-time performance" on resource-constrained devices.

[0049] 2. Complementary advantages of hybrid architecture: The hybrid architecture of "pre-set resource library + on-demand generation engine" ensures the reliability and consistency of content such as accurate scientific models, while meeting users' personalized creation needs through the generation engine, thus balancing accuracy, efficiency and flexibility.

[0050] 3. Possesses adaptive optimization capabilities: Through its built-in learning and optimization module, the system can continuously learn user preferences and adapt to changes in device status, automatically optimize scheduling strategy parameters, and has the ability to evolve over a long period of time, demonstrating a high degree of intelligence.

[0051] 4. System decoupling and strong versatility: The method and system provided by this invention adopt a modular design, with each module having a clear function and interface. It does not depend on a specific hardware carrier, has strong scalability, and is easy to integrate into various types of resource-constrained intelligent interactive devices. Attached Figure Description

[0052] Figure 1 A flowchart illustrating a method for generating and scheduling hybrid content for resource-constrained devices;

[0053] Figure 2 A schematic diagram illustrating the decision logic of the strategy scheduling algorithm for generating scheduling methods for hybrid content;

[0054] Figure 3 A structural diagram of a scheduling system for hybrid content generated for resource-constrained devices; Detailed Implementation

[0055] The present invention will be further illustrated below with reference to specific embodiments, but the embodiments do not limit the present invention in any way. Unless otherwise specified, the reagents, methods, and equipment used in the present invention are conventional reagents, methods, and equipment in this technical field.

[0056] In one embodiment of the present invention, a hybrid content generation and scheduling method for resource-constrained devices is provided, such as... Figure 1 As shown, it includes the following steps:

[0057] S1. Receive user content generation requests, parse the requests, and obtain the request type, content topic, and complexity score; the request type is simple query, complex creation, or precise generation.

[0058] S2. Based on the parsing results, evaluate the current resource status of the system, call the strategy scheduling algorithm, and dynamically select the path between the preset resource library and the on-demand generation engine.

[0059] S3. If the strategy scheduling algorithm decides to call the preset resource library, then perform multi-level matching and retrieval based on the content topic and output the optimal preset content.

[0060] If the strategy scheduling algorithm decides to activate the on-demand generation engine, it will allocate computing resources and call a lightweight generation model or cloud service to generate the target content based on the content theme and complexity score.

[0061] S4. Render and optimize the optimal preset content or target content and adapt its format, then output it to the terminal display module.

[0062] In step S1, when parsing the request, a natural language processing model is used to parse the request and extract the key instruction entities in the request.

[0063] Determine the request type based on a predefined rule base;

[0064] Based on the abstraction level of the instruction entity and the required computational resources, a scoring model is used to quantify the complexity. The expression of the scoring model is:

[0065] ;

[0066] Where C is the complexity score, L is the abstraction level quantification value of the instruction entity, R is the estimated computational resources required, and w1 and w2 are preset weight coefficients.

[0067] In step S2, the decision logic of the strategy scheduling algorithm is based on a multidimensional evaluation matrix. The input variables of the multidimensional evaluation matrix include: the complexity score of the request; the matching confidence of the content in the preset resource library; the current available computing resources of the system (CPU / GPU utilization, memory availability); network bandwidth and latency status (used to evaluate the feasibility of cloud generation); the estimated generation time and power consumption of the on-demand generation engine; and the output variable is a decision path that can be chosen from two options.

[0068] The policy scheduling algorithm employs a threshold-based decision tree model, such as... Figure 2 As shown, its implementation method is as follows:

[0069] If the request type is determined to be "exactly generated" or the matching confidence is higher than the first set threshold (Th_predefined), then the default resource library will be called first.

[0070] If the request type is determined to be "complex creation" or the matching confidence is lower than the second set threshold (Th_generate), it is recommended to call the preset resource library;

[0071] If the request type is determined to be complex creation or the matching confidence is higher than the second set threshold (Th_generate), and the system's current available resources are higher than the resource security threshold (Th_resource), then the on-demand generation engine will be activated first.

[0072] If the system's current available resources are below the resource safety threshold (Th_resource), a forced decision will be made to call the preset resource library or return a resource shortage waiting instruction.

[0073] In step S3, multi-level matching and retrieval include:

[0074] First-level matching: Matches content themes (such as "chemistry", "biology", "art style") to first-level category nodes;

[0075] Second-level matching: Calculate vector similarity under the second-level category node based on the requested attribute tags (such as "water molecule" or "cartoon");

[0076] Search optimization: Based on the collaborative filtering algorithm, it recommends preset content with high relevance according to historical interaction data.

[0077] In step S3, the on-demand generation engine has a modular structure, including:

[0078] The image generation unit generates lightweight 3D virtual images based on text descriptions.

[0079] The formula parsing unit converts abstract information such as mathematical formulas and chemical molecular formulas into three-dimensional geometric structures that can be rendered.

[0080] The style transfer unit transforms the artistic style of existing preset content;

[0081] The generation process adopts a "cloud-edge" collaborative computing framework. Lightweight models run on local devices, while large models call cloud APIs and perform localized post-processing on the generated results.

[0082] In step S4, rendering optimization and format adaptation include:

[0083] The model is lightweighted, reducing the number of faces and texture resolution to a level that allows for smooth rendering on the target screen;

[0084] Material and lighting effects are remapped to adapt the rendering pipeline to specific display hardware;

[0085] The orientation and scale of the content are automatically adjusted based on the optimal viewing angle of the display device.

[0086] This embodiment also provides a hybrid content generation scheduling system for resource-constrained devices, used to implement the above-described hybrid content generation scheduling method for resource-constrained devices. The hybrid content generation scheduling system is as follows: Figure 3 As shown, it includes:

[0087] The request parsing module is used to receive and parse the user's content generation request, and output the request type, content topic, and complexity score;

[0088] The strategy scheduling module, which has an embedded strategy scheduling algorithm, is used to make content generation path decisions based on the output of the request parsing module.

[0089] The preset resource library module is used to store, index, and manage preset 3D models, animations, and special effects content, and to perform retrieval tasks;

[0090] The on-demand generation engine module is used to generate target content by calling local or cloud models after receiving instructions;

[0091] The rendering adaptation module is used to optimize the generated or retrieved content to adapt to the output requirements of the terminal display module.

[0092] The terminal display module is used to display the generated content.

[0093] The hybrid content generation system also includes a learning optimization module, which records each content request and its corresponding path decision and result feedback; based on the recorded data, it uses reinforcement learning algorithms to iteratively optimize the decision threshold and weight parameters in the strategy scheduling algorithm.

[0094] In one specific embodiment of the present invention, taking an intelligent interactive device integrating light field display technology as an example, the device typically has the characteristics of limited computing power, memory and power consumption, and is suitable for interactive scenarios such as education and display that require high real-time and high-precision content generation.

[0095] Users can issue commands via voice or text input, such as: "Please show a water molecule model" and "Please convert it to a cartoon rendering style".

[0096] S1. Receive user content generation request: The request parsing module performs entity recognition and intent classification using the Natural Language Processing (NLP) model integrated into the device.

[0097] "Water molecule model": Identified as a "chemical molecule" entity, intended to be classified as "precisely generated", with a complexity rating of low.

[0098] "Cartoon Rendering Style": Identified as a "Visual Style" entity, intended to be classified as "Complex Creation", requiring content generation and transformation, with a complexity rating of medium.

[0099] Step S2, Strategy Scheduling Decision:

[0100] The policy scheduling module makes decisions based on the comprehensive analysis results (CPU utilization, memory availability), and its decision logic is based on a multi-dimensional evaluation matrix. The input variables are normalized and weighted before comprehensive evaluation.

[0101] For the "water molecule model": A query of the preset resource library revealed a high-precision H2O model with a matching confidence level >99%, and the request was for "exact generation". Algorithm decision: Call the preset resource library.

[0102] For the "cartoon rendering style": The default resource library is queried, and the confidence level for matching related styles is low, while the current system computing power is sufficient. Algorithm decision: Activate the on-demand generation engine (style transfer unit). If system resource monitoring shows that available memory is below the safe threshold (Th_resource), the system will activate a degradation strategy, forcibly selecting and calling the closest existing stylization model in the default resource library to ensure system stability and real-time performance.

[0103] Step S3: Call the preset resource library:

[0104] The preset resource library module, under the path of "Chemistry / Molecular Models / Common Inorganic Substances", retrieves and outputs a high-precision 3D model of the H2O molecule through multi-level indexing and vector similarity matching.

[0105] Launch the on-demand generation engine:

[0106] The style transfer unit of the on-demand generation engine is activated. It loads the base "water molecule" 3D model and then calls the locally running lightweight style transfer model to generate a "cartoon-rendered" water molecule model.

[0107] Step S4, Display Adaptation and Output:

[0108] The rendering adaptation module optimizes the output 3D model, including adjusting the level of detail (LOD) based on screen parameters and remapping the lighting and shadow effects, and then sends it to the light field display module for stereoscopic rendering output.

[0109] The timing of data interaction in the system of this invention is as follows:

[0110] 1. The user sends a content generation request to the request parsing module;

[0111] 2. After the request parsing module completes the parsing, it sends the structured request information (request type, content topic, complexity score) to the strategy scheduling module;

[0112] 3. The strategy scheduling module queries the real-time system status;

[0113] 4. The strategy scheduling module makes path decisions based on the parsing results and resource status;

[0114] 5. If the decision is to call the preset resource library, a search command is sent to the preset resource library module, which returns the preset content;

[0115] 6. If the decision is to initiate generation, a generation command is sent to the on-demand generation engine module, which generates the content and returns it.

[0116] 7. The preset resource library module or the on-demand generation engine module sends the acquired content to the rendering adaptation module;

[0117] 8. After the rendering and adaptation module completes optimization, the final content is sent to the display device for output;

[0118] 9. The learning optimization module asynchronously records the log data of the entire interaction process for subsequent strategy optimization.

[0119] This timing sequence confirms the workflow and data flow of the collaborative work between the modules of the system described in this invention, reflecting the dynamism and real-time nature of the scheduling decision.

[0120] The present invention provides a hybrid content generation and scheduling method and system for resource-constrained devices, which enables the generation of high-quality, real-time and personalized intelligent content on resource-constrained embedded devices, providing core technical support for the next generation of human-computer interaction devices.

[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for hybrid content generation and scheduling for resource-constrained devices, characterized in that, Includes the following steps: S1. Receive user content generation requests, parse the requests, and obtain the request type, content topic, and complexity score; the request type is simple query, complex creation, or precise generation. S2. Based on the parsing results, evaluate the current resource status of the system, call the strategy scheduling algorithm, and dynamically select the path between the preset resource library and the on-demand generation engine. The decision logic of the strategy scheduling algorithm is based on a multidimensional evaluation matrix. The input variables of the multidimensional evaluation matrix include: the complexity score of the request; the matching confidence of the request content in the preset resource library; the current available computing resources of the system; network bandwidth and latency status; the estimated generation time and power consumption of the on-demand generation engine; and the output variable is a decision path with two choices. The strategy scheduling algorithm adopts a threshold-based decision tree model. S3. If the strategy scheduling algorithm decides to call the preset resource library, then perform multi-level matching and retrieval based on the content topic and output the optimal preset content. If the strategy scheduling algorithm decides to activate the on-demand generation engine, it will allocate computing resources and call the generation model to generate the target content based on the content topic and complexity score. S4. Render and optimize the optimal preset content or target content and adapt its format, then output it to the terminal display module.

2. The hybrid content generation and scheduling method for resource-constrained devices according to claim 1, characterized in that, In step S1, when parsing the request, a natural language processing model is used to parse the request and extract the key instruction entities in the request; Determine the request type based on a predefined rule base; Based on the abstraction level of the instruction entity and the required computing resources, a scoring model is used to quantify the complexity score.

3. The hybrid content generation and scheduling method for resource-constrained devices according to claim 2, characterized in that, The expression for the scoring model is: ； Where C is the complexity score, L is the abstraction level quantification value of the instruction entity, R is the estimated computational resources required, and w1 and w2 are preset weight coefficients.

4. The hybrid content generation and scheduling method for resource-constrained devices according to claim 1, characterized in that, The implementation method of the policy scheduling algorithm is as follows: If the request type is determined to be exactly generated or the matching confidence level is higher than the first set threshold, the default resource library will be called first. If the request type is determined to be complex creation or the matching confidence is lower than the second set threshold, then the preset resource library will be called. If the request type is determined to be complex creation or the matching confidence is higher than the second set threshold, and the system's current available resources are higher than the resource safety threshold, then the on-demand generation engine will be activated first. If the system's current available resources are below the resource safety threshold, a forced decision will be made to call the preset resource library or return a resource shortage waiting instruction.

5. The hybrid content generation and scheduling method for resource-constrained devices according to claim 1, characterized in that, In step S3, the multi-level matching and retrieval includes: First-level matching: Matches content to the first-level category node based on the content theme; Second-level matching: Calculate vector similarity under the second-level category node based on the requested attribute tags; Search optimization: Based on the collaborative filtering algorithm, it recommends preset content with high relevance according to historical interaction data.

6. The hybrid content generation and scheduling method for resource-constrained devices according to claim 1, characterized in that, In step S3, the on-demand generation engine has a modular structure, including: The image generation unit generates lightweight 3D virtual images based on text descriptions. The formula parsing unit converts abstract information into a three-dimensional geometric structure that can be rendered; The style transfer unit transforms the artistic style of existing preset content; The generation process adopts a cloud-edge collaborative computing framework. Lightweight models run on local devices, while large models call cloud APIs and perform localized post-processing on the generated results.

7. The hybrid content generation and scheduling method for resource-constrained devices according to claim 1, characterized in that, In step S4, the rendering optimization and format adaptation include: The model is lightweighted, reducing the number of faces and texture resolution to a level that allows for smooth rendering on the target screen; Material and lighting effects are remapped to adapt the rendering pipeline to specific display hardware; The orientation and scale of the content are automatically adjusted based on the optimal viewing angle of the display device.

8. A hybrid content generation and scheduling system for resource-constrained devices, characterized in that, For implementing the hybrid content generation scheduling method for resource-constrained devices according to any one of claims 1-7, the hybrid content generation scheduling system comprises: The request parsing module is used to receive and parse the user's content generation request, and output the request type, content topic, and complexity score; The strategy scheduling module, which incorporates the strategy scheduling algorithm, is used to make content generation path decisions based on the request parsing module. The preset resource library module is used to store, index, and manage preset 3D models, animations, and special effects content, and to perform retrieval tasks; The on-demand generation engine module is used to generate target content by calling local or cloud models after receiving instructions; The rendering adaptation module is used to optimize the generated or retrieved content to adapt to the output requirements of the terminal display module. The terminal display module is used to display the generated content.

9. The hybrid content generation and scheduling system for resource-constrained devices according to claim 8, characterized in that, It also includes a learning optimization module, which records each content request and its corresponding path decision and result feedback; based on the recorded data, it uses a reinforcement learning algorithm to iteratively optimize the decision threshold and weight parameters in the policy scheduling algorithm.

Images