A video generation service system for large language models supporting autonomous intelligent agents

By using engine capability detection, missing item list maintenance, and description replacement, the problem of resource waste in autonomous intelligent agent video generation is solved, and efficient video generation task management is achieved.

CN123123673APending Publication Date: 2026-11-17BEIJING VECTOR EQUATION TECHNOLOGY CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611405368.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, in autonomous intelligent agent video generation scenarios, it is impossible to determine whether the task can be completed before generation, resulting in wasted generation resources, and if the verification fails, it can only be regenerated.

Method used

A capability detection module is introduced to detect engines, a profile registration module maintains a list of missing items, a pre-interception module filters engines, an expression replacement module provides expression fault tolerance, and a generation scheduling module optimizes the generation of tasks.

Benefits of technology

By using engine capabilities for filtering and fault tolerance, invalid generation is avoided, resource waste is reduced, and the executability and resource utilization efficiency of generation tasks are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123673A_ABST
    Figure CN123123673A_ABST
Patent Text Reader

Abstract

This invention relates to the field of video generation, and more particularly to a large language model video generation service system supporting autonomous agents. The system schedules video generation engines to execute a set of detection tasks when detection trigger conditions are met, performs item detection on the generated results, and registers missing items as a list of omissions. The video generation tasks submitted by the autonomous agent are identified as a list of essential content. Before generation scheduling, the list of essential content is compared with the omission lists of each candidate engine. Engines whose omission lists contain items from the essential content list are excluded. If all are excluded, the omissions are replaced using an expression replacement lookup table, and the comparison is repeated. If this is still not feasible, the system is prohibited from entering the generation scheduling process, and engines that have not been excluded are scheduled to generate the target video. This invention, through pre-emptive engine capability screening and expression fault tolerance, only schedules engines capable of completing the essential content to perform generation, avoiding the consumption of computing power by invalid generation and reducing the waste of generation resources.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of video generation, and more particularly to a large language model video generation service system that supports autonomous intelligent agents. Background Technology

[0002] With the development of large language models and video generation technology, video generation services are gradually transforming from creation tools for manual operation to basic services for autonomous intelligent agents. Autonomous agents autonomously submit video generation tasks based on task planning, obtain generation results, and execute subsequent processes. The automated acceptance and delivery of video generation tasks has become a core requirement for the construction of generation service platforms. Existing video generation services typically adopt a service model where the engine is directly scheduled to generate the video after receiving the task, and then the quality of the generated results is verified. If the verification fails, generation is re-executed. The scheduling stage often selects the generation engine based on cost and resource availability. However, video generation tasks have high computational overhead and long generation times. Each verification failure means that all generation computing power has been consumed. Furthermore, different video generation engines have systematic differences in their ability to complete different types of generated content. Before task execution, it is impossible to predict the risk of missing specific content in the selected engine for the current task, and the adaptation experience accumulated after engine version updates becomes invalid. These problems result in a large number of invalid generation tasks in video generation services, and the existing generation and verification mechanisms are still insufficient in dealing with the waste of generation resources and the unpredictability of engine capabilities.

[0003] Chinese Patent Publication No. CN120378710B discloses a video generation method, system, device, and medium based on a large language model. The method involves: acquiring product information input by a user, including the product's name, description, and selling points; preprocessing the product information and then performing semantic segmentation using a large language model to obtain segmentation description information corresponding to the product information; acquiring the original video and segmenting it into multiple video segments and corresponding scene description information; performing semantic matching on the segmentation description information and scene description information to obtain the video segment with the highest similarity to each segmentation description information; and concatenating the video segments according to the order of the segmentation description information to generate a complete video. Compared to related technologies, this application, by integrating natural language processing, video content analysis, and intelligent matching algorithms, can automatically and efficiently generate product videos.

[0004] However, the following problems still exist in the existing technology: In existing technologies, the content that each generation engine struggles to complete is not recorded. Before generation, it is impossible to determine whether the selected engine can complete the current task. If the verification fails, the engine can only be regenerated. Summary of the Invention

[0005] To address this issue, the present invention provides a large language model video generation service system that supports autonomous intelligent agents, thereby overcoming the problem in the prior art where, in autonomous intelligent agent video generation scenarios, it is impossible to determine whether the task can be completed before generation, and if it fails, it can only be regenerated, resulting in wasted generation resources.

[0006] To achieve the above objectives, the present invention provides a large language model video generation service system supporting autonomous intelligent agents, comprising: The capability detection module is used to schedule the video generation engine to execute a preset set of detection tasks in response to the fulfillment of preset detection trigger conditions, so as to determine the generation results of each detection. The image registration module is used to perform item detection on the generated results to filter out items that do not appear in the generated results from the detection task set, and register the filtered items as a list of missing items in the video generation engine. The task receiving module is used to receive video generation tasks submitted by autonomous intelligent agents and determine a list of essential content based on the video generation tasks. The pre-interception module is used to compare the items in the essential content list with the omission list of each candidate video generation engine before the video generation task enters the generation scheduling. If the omission list of a candidate video generation engine contains an item in the essential content list, the candidate video generation engine is excluded. The expression replacement module is used in response to the exclusion of all candidate video generation engines. Based on the preset expression replacement comparison table, it replaces the items in the essential content list that are the same as the missing items list with the corresponding equivalent expression items, and re-executes the comparison based on the replaced essential content list. If there are still no candidate video generation engines that have not been excluded after replacement, the video generation task is prohibited from entering the generation schedule. The generation scheduling module is used to schedule the candidate video generation engines that have not been excluded to perform video generation tasks in order to generate the target video.

[0007] Furthermore, the process by which the capability detection module determines whether the preset detection triggering conditions are met includes: Monitor the access status and engine identifier of the video generation engine; If the access of the video generation engine is detected or the engine identifier of the video generation engine is changed, it is determined that the preset detection triggering conditions are met.

[0008] Furthermore, the portrait registration module includes: An entry detection submodule is used to perform entry detection on the generated result to determine the occurrence of each entry in the detection task set in the generated result; The missing item filtering submodule is used to filter items that do not appear in the generated results based on their occurrence. The list registration submodule is used to register the item as a missing item list of the video generation engine according to the category code when the number of consecutive absences of an item reaches a preset registration number. The omission list is stored in association with the engine identifier of the video generation engine.

[0009] Furthermore, the task receiving module includes: The task parsing submodule is used to perform content parsing on the video generation task to determine the content items that the video generation task requires to be presented; The coding and labeling submodule is used to assign category codes to content items based on the category coding system and organize the content items with assigned category codes into a list of essential content. The necessity labeling submodule is used to determine the necessity of each item in the essential content list; The necessity indicator includes a necessity level and a desired level.

[0010] Furthermore, the pre-interception module includes: The encoding extraction submodule is used to extract the category codes of each item in the essential content list and extract the category codes of each item in the missing item list of each candidate video generation engine. The matching and determination submodule is used to determine that items with the same category code are the same items, and based on the determination result of the same items, it determines whether the missing items list contains items from the required content list. The items in the essential content list and the missing items list are determined based on the same category coding system.

[0011] Furthermore, the items in the essential content list carry necessity markers; the process by which the pre-interception module excludes candidate video generation engines includes: If the omission list of candidate video generation engines contains items marked as necessary, then the candidate video generation engine is excluded. If the list of missing items in the candidate video generation engine contains items marked as "desired" in terms of necessity, then the desired level hit rate of the candidate video generation engine is recorded, and the candidate video generation engine is retained.

[0012] Furthermore, the process by which the generation scheduling module schedules the candidate video generation engines that have not been excluded includes: Obtain the expected hit rate of candidate video generation engines that have not been excluded; The candidate video generation engine with the fewest expected hits is selected to perform the video generation task.

[0013] Furthermore, the expression replacement module includes: The replacement query submodule is used to identify items in the required content list that are the same as those in the missing items list as items to be replaced, and to determine the replacement rule corresponding to the item to be replaced based on a preset expression replacement lookup table; the expression replacement lookup table stores the correspondence between item types and replacement rules; The representation generation submodule is used to generate equivalent representation entries corresponding to the entries to be replaced according to the replacement rules, and write the equivalent representation entries back to the required content list; The equivalent representation entry inherits the category code of the entry to be replaced.

[0014] Furthermore, the expression replacement module prevents video generation tasks from entering the generation schedule process, including: Record the number of replacements performed on the essential content list; If the number of replacements reaches the preset replacement limit, and there are still no candidate video generation engines that have not been excluded after replacement, then the video generation task is prohibited from entering the generation schedule, an infeasibility receipt is generated, and the infeasibility receipt is returned to the autonomous intelligent agent. The infeasibility receipt contains information about the same items in the required list as those in the missing items list.

[0015] Furthermore, the scheduling generation module includes: The generation execution submodule is used to schedule the candidate video generation engines that have not been excluded to perform video generation tasks in order to generate the target video; The finished product inspection submodule is used to perform item inspection on the target video and filter out items from the list of essential content that do not appear in the target video. The segment regeneration submodule is used to determine the time period corresponding to the selected items in the video generation task, and to schedule the candidate video generation engine to regenerate the corresponding video segment for the time period.

[0016] Compared with existing technologies, this invention schedules video generation engines to execute a set of detection tasks when the detection trigger conditions are met, performs item detection on the generated results, and registers missing items as a list of omissions; it also identifies the video generation tasks submitted by the autonomous agent as a list of essential content; before generation scheduling, it compares the list of essential content with the omission lists of each candidate engine, excluding engines whose omission lists contain items from the list of essential content. If all are excluded, the engine is replaced according to the expression replacement lookup table and re-compared. If it is still not feasible, it is prohibited from entering the generation scheduling process, and the engine that has not been excluded is scheduled to generate the target video. This invention, through pre-emptive engine capability screening and expression fault tolerance, only schedules engines capable of completing the essential content to perform generation, avoiding the consumption of computing power by invalid generation and reducing the waste of generation resources.

[0017] In particular, this invention dynamically maintains a list of missing items for each video generation engine through a capability detection module and a profile registration module, enabling continuous profiling of the engine's generation capabilities. Different video generation engines exhibit systematic characteristics in their missing specific content items, and the adaptation experience accumulated after engine version updates becomes invalid. When this invention detects engine access or a change in engine identifier, it schedules the engine to execute a pre-set set of detection tasks, performs item detection on the generated results, and registers items that are missing in multiple consecutive detections as a missing item list according to category codes, storing them in association with the engine identifier. This missing item list provides a quantitative basis for selecting available engines before subsequent task execution, avoiding the computational waste caused by blindly scheduling engines with unknown capabilities.

[0018] In particular, this invention introduces a pre-interception mechanism before generation scheduling, and combines necessity flags to distinguish the different levels of importance of items in the essential content list, thereby filtering candidate engines. The content items required for video generation tasks can be divided into necessary items that must be presented and expected items that are expected to be presented. If a candidate engine is directly excluded due to the absence of expected items, available engine resources will be lost. This invention, through a pre-interception module, compares the essential content list with the omission lists of each candidate engine based on the same category coding system. Only when the omission list contains items with the necessity flag set to "necessary" is the candidate engine excluded; if only expected items are matched, the expected-level match is recorded and the engine is retained. This invention maximizes the utilization of the candidate engine's generation capabilities and reduces the possibility of misjudging available engines due to non-critical items.

[0019] In particular, this invention solves the problem of all engines being excluded due to semantic expression differences by using an expression replacement module, further improving the executability of video generation tasks. The same content item can often be presented in multiple equivalent expressions, and different video generation engines have different response capabilities to different expressions of the same semantic meaning. If the expression of an item in the essential content list does not match the intermediate representation features of the engine, even if the engine actually has the ability to generate the content, it may still be incorrectly excluded because the omission list contains that item. In this invention, after all candidate engines are excluded during pre-interception, the expression replacement module is invoked to query a pre-set expression replacement lookup table, converting items to be replaced that conflict with the omission list into semantically equivalent expression items and inheriting their category codes. Then, based on the replaced essential content list, a comparison with the omission lists of each candidate engine is re-executed. If a candidate engine that has not been excluded is found, that engine is scheduled to generate the content. Only when no engine is available after the number of replacements reaches a preset limit is an infeasibility receipt generated, and the task is prohibited from entering the generation schedule. This invention effectively expands the applicability of candidate engines, avoids a large number of tasks that could have been completed failing directly due to expression differences, and reduces the waste of generation resources.

[0020] In particular, this invention further refines the repair of occasional missing essential items in the generated results through a finished product detection and segment regeneration mechanism in the generation scheduling module. Even if high-risk engines are accurately excluded through pre-interception, the video generation engine may still experience occasional missing essential items in a single generation. Existing solutions typically require regenerating the entire video after a finished product verification failure, resulting in wasted computing power. This invention performs item detection on the generated target video, filters out essential content list items that do not appear, determines the time period corresponding to the item in the video generation task, and the segment regeneration submodule schedules the same candidate engine to regenerate the corresponding video segment only for that time period. This invention reduces the overhead of repeated generation after verification failure and further improves the resource utilization efficiency of the generation service. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of a large language model video generation service system supporting autonomous intelligent agents according to an embodiment of the present invention; Figure 2 This is a flowchart of the omission list registration process according to an embodiment of the present invention; Figure 3 This is a logic diagram for pre-interception in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the correspondence of the substitution lookup table in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the operation of a large language model video generation service system supporting autonomous intelligent agents, as described in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] Please see Figure 1As shown, this is a schematic diagram of the structure of a large language model video generation service system supporting autonomous intelligent agents according to an embodiment of the present invention. The system of this embodiment includes: The capability detection module is used to schedule the video generation engine to execute a preset set of detection tasks in response to the fulfillment of preset detection trigger conditions, so as to determine the generation results of each detection. The image registration module is used to perform item detection on the generated results to filter out items that do not appear in the generated results from the detection task set, and register the filtered items as a list of missing items in the video generation engine. The task receiving module is used to receive video generation tasks submitted by autonomous intelligent agents and determine a list of essential content based on the video generation tasks. The pre-interception module is used to compare the items in the essential content list with the omission list of each candidate video generation engine before the video generation task enters the generation scheduling. If the omission list of a candidate video generation engine contains an item in the essential content list, the candidate video generation engine is excluded. The expression replacement module is used in response to the exclusion of all candidate video generation engines. Based on the preset expression replacement comparison table, it replaces the items in the essential content list that are the same as the missing items list with the corresponding equivalent expression items, and re-executes the comparison based on the replaced essential content list. If there are still no candidate video generation engines that have not been excluded after replacement, the video generation task is prohibited from entering the generation schedule. The generation scheduling module is used to schedule the candidate video generation engines that have not been excluded to perform video generation tasks in order to generate the target video.

[0026] Specifically, the overall solution of this invention automatically executes detection tasks and dynamically maintains a list of missing items when an engine is accessed or changed through a capability detection module and a profile registration module, thus constructing a quantifiable capability boundary description for each video generation engine. The task receiving module extracts a list of essential content from the video generation tasks submitted by the autonomous agent and marks its necessity level. Before the task enters the generation scheduling stage, the pre-interception module compares the list of essential content with the list of missing items of each candidate engine at the category coding level, excluding high-risk engines and retaining engines with expected hits based on necessity markers, thereby completing the screening of available engines before computing power is invested. When all candidate engines are excluded because items in their essential content lists conflict with their missing item lists, the expression replacement module converts the conflicting items into equivalent expressions based on a pre-set expression replacement lookup table and re-executes the comparison under the constraint of the number of replacements, expanding the applicability of candidate engines through an expression fault tolerance mechanism. The generation scheduling module then selects the engine with the fewest expected missing items from the selected available engines for generation, and after generation, through finished product detection and segment regeneration mechanisms, only locally regenerates the time periods of individual missing items. This invention reduces invalid generation caused by insufficient engine capabilities, and replaces complete regeneration with fragment-level repair, thereby reducing the recalculation cost of failed tasks and achieving efficient utilization of generation computing power overall.

[0027] Specifically, the video generation engine refers to a video generation channel that generates video synthesis code from a large language model and then renders the video synthesis code into a video by a rendering engine. The video synthesis code is structured readable text. Each frame of the target video is determined by the rendering engine evaluating the video synthesis code over time, and multiple renderings of the same video synthesis code yield consistent results. The system of this embodiment maintains several video generation engines simultaneously, and each video generation engine uses different intermediate representations and content expression capabilities.

[0028] In this embodiment, several video generation engines include an engine that uses a component tree and declarative time slices as intermediate representations, an engine that uses a page document and a program-controllable pause timeline as intermediate representations, and a hand-drawn engine that uses vector path data with drawing order and stroke weights as intermediate representations. The capabilities of each video generation engine for completing specific content categories vary systematically. For example, only some video generation engines support the overlay of live-action narration content, while the hand-drawn engine presents illustration content through a stroke-by-stroke drawing process. After the video generation engine is updated, the missing item list corresponding to the original version is no longer applicable to the updated video generation engine.

[0029] It is understood that there are no restrictions on the specific implementation of the video generation engine. It can be constructed using any intermediate representation and rendering pipeline, as long as it can respond to the video generation task and produce video, and the produced video can be used for entry detection. Further details will not be provided here.

[0030] It is understood that the video generation service system provided by this invention is particularly suitable for automated video production platforms for autonomous intelligent agents. Autonomous intelligent agents submit video generation tasks in batches according to task planning, without needing to be aware of the capability differences and version changes of the underlying video generation engines; the system can automatically complete engine adaptation and generation scheduling. Based on this, the technical solution of this system can be extended to other multimodal content generation scenarios where differences in engine capabilities lead to wasted generation resources, such as image generation, 3D model rendering, or speech synthesis, etc., which will not be elaborated further here.

[0031] It should be noted that the multiple functional modules involved in this application are only logical divisions based on the functions implemented according to the present invention. The specific physical structures of the capability detection module, profile registration module, task receiving module, pre-interception module, expression replacement module, and generation scheduling module are not limited. They can be composed of logical components, including field-programmable gate arrays, application-specific integrated circuits, computers, or microprocessors in computers, which will not be elaborated here.

[0032] Please see Figure 5 The diagram shown is a flowchart illustrating the operation of a large language model video generation service system supporting autonomous intelligent agents, according to an embodiment of the present invention. The system operation process is as follows: In step S1, the capability detection module responds to the preset detection triggering conditions by scheduling the video generation engine to execute a preset set of detection tasks. The profile registration module performs item detection on the generation results of the detection and registers the consecutive items that do not appear in the generation results as the omission list of the video generation engine. Step S2: The task receiving module receives the video generation task submitted by the autonomous intelligent agent and determines a list of essential content carrying necessity flags based on the video generation task. Step S3: Before the video generation task enters the generation schedule, the pre-interception module compares the list of essential content with the list of missing items of each candidate video generation engine by category code, excludes candidate video generation engines whose list of missing items contains necessary items, and records the expected hit status of candidate video generation engines whose list of missing items only contains expected items and retains them. In step S4, in response to the exclusion of all candidate video generation engines, the expression replacement module replaces the items to be replaced in the essential content list with equivalent expression items that inherit the category code based on the preset expression replacement comparison table, and re-executes the comparison; if the number of replacements reaches the preset replacement limit and there are still no candidate video generation engines that have not been excluded, the video generation task is prohibited from entering the generation schedule, and an infeasibility receipt is returned to the autonomous intelligent agent. Step S5: The scheduling module selects the candidate video generation engine with the fewest expected hits from the candidate video generation engines that have not been excluded and executes the video generation task to generate the target video. Step S6: The scheduling module performs item detection on the target video, filters out items from the essential content list that do not appear in the target video, and regenerates the corresponding video segments for the time periods corresponding to the filtered items to obtain the final target video.

[0033] Specifically, the process by which the capability detection module determines whether the preset detection trigger conditions are met includes: Monitor the access status and engine identifier of the video generation engine; If the access of the video generation engine is detected or the engine identifier of the video generation engine is changed, it is determined that the preset detection triggering conditions are met.

[0034] In this embodiment, the system performs a gray-scale switching of several video generation engines: a unified service entry point distributes video generation tasks to each video generation engine according to routing decisions, and the same user is always routed to the same video generation engine. The engine identifier of the hit video generation engine is recorded in the task record; the routing decision supports percentage-based volume, whitelist targeting, and circuit breaker switching back; if a new video generation engine is detected to be connected, or if the engine identifier of an existing video generation engine is detected to be changed, it is determined that the preset detection triggering conditions are met, and a detection task set is executed on the video generation engine to reconstruct the missing item list of the video generation engine.

[0035] Specifically, the portrait registration module includes: An entry detection submodule is used to perform entry detection on the generated result to determine the occurrence of each entry in the detection task set in the generated result; The missing item filtering submodule is used to filter items that do not appear in the generated results based on their occurrence. The list registration submodule is used to register the item as a missing item list of the video generation engine according to the category code when the number of consecutive absences of an item reaches a preset registration number. The omission list is stored in association with the engine identifier of the video generation engine.

[0036] Specifically, the omission list refers to a collection of content items that the video generation engine has difficulty completing. The items in the omission list are organized by category code and stored in association with the engine identifier of the video generation engine. The omission list is used to screen candidate video generation engines that do not have the necessary content generation capabilities before the video generation task enters the generation scheduling.

[0037] Specifically, the detection task set is organized by category code. The content categories corresponding to the category code include the existence, quantity, color, spatial relationship, material, camera movement, temporal action, and the image of a live narrator. Each detection task carries several items to be verified. Multiple detection tasks are set up under the same content category to support continuous verification of the same item.

[0038] In this embodiment, the item detection is performed by the content understanding unit. The content understanding unit extracts video frames from the generated results at several time points. For each item to be verified carried by the detection task, the unit gives a judgment result of whether the item has appeared or not for each extracted video frame. The detection task is scheduled to be executed when the system load is lower than a preset level, without occupying online generation resources.

[0039] Please see Figure 2 The process for registering the omission list is as follows: Monitor the engine access status and engine identifier, and schedule the engine to execute a set of detection tasks when the detection trigger conditions are met; The generated results are checked for entries. When an entry fails to appear for a certain number of consecutive times, the corresponding entry is registered as a missing item list according to the category code and stored in association with the engine identifier.

[0040] It is understandable that the specific implementation method of item detection is not limited. Any means that can determine whether an item appears or not in the generated result can be used. The value of the number of registrations can be configured according to the engine stability requirements. The larger the value, the fewer false judgments there will be in the missing item list, and the more detection rounds are required for registration. This will not be elaborated further.

[0041] Specifically, the task receiving module includes: The task parsing submodule is used to perform content parsing on the video generation task to determine the content items that the video generation task requires to be presented; The coding and labeling submodule is used to assign category codes to content items based on the category coding system and organize the content items with assigned category codes into a list of essential content. The necessity labeling submodule is used to determine the necessity of each item in the essential content list; The necessity indicator includes a necessity level and a desired level.

[0042] Specifically, the essential content list refers to the set of content items required to be presented in the video generation task. Each item in the essential content list is assigned a category code and a necessity indicator.

[0043] In this embodiment, the autonomous intelligent agent submits a video generation task through a unified service portal based on its own task plan. The video generation task carries the target duration, content requirements, and materials. The task parsing submodule performs content parsing on the content requirements to determine the content items that the video generation task requires to be presented. The necessity labeling submodule determines the content items corresponding to the materials that must be used as necessary level and the content items used to enhance the picture performance as expected level.

[0044] Specifically, the pre-interception module includes: The encoding extraction submodule is used to extract the category codes of each item in the essential content list and extract the category codes of each item in the missing item list of each candidate video generation engine. The matching and determination submodule is used to determine that items with the same category code are the same items, and based on the determination result of the same items, it determines whether the missing items list contains items from the required content list. The items in the essential content list and the missing items list are determined based on the same category coding system.

[0045] Specifically, the items in the essential content list carry a necessity indicator; the process by which the pre-interception module excludes candidate video generation engines includes: If the omission list of candidate video generation engines contains items marked as necessary, then the candidate video generation engine is excluded. If the list of missing items in the candidate video generation engine contains items marked as "desired" in terms of necessity, then the desired level hit rate of the candidate video generation engine is recorded, and the candidate video generation engine is retained.

[0046] Specifically, the process by which the generation scheduling module schedules candidate video generation engines that have not been excluded includes: Obtain the expected hit rate of candidate video generation engines that have not been excluded; The candidate video generation engine with the fewest expected hits is selected to perform the video generation task.

[0047] Specifically, the expected hit record contains the correspondence between the retained candidate video generation engines and the expected items in the required content list, including the number of expected items hit and their category codes.

[0048] Please see Figure 3 The judgment process for pre-interception is as follows: After comparing the list of required content with the list of missing items of each candidate engine, candidate engines that contain necessary items in their missing items list are excluded, candidate engines that do not contain necessary items are retained, and the expected hit rate is recorded.

[0049] In this embodiment, the generation scheduling module obtains the expected level hit status of each candidate video generation engine that has not been excluded, counts and compares the hit expected level entries, and selects the candidate video generation engine with the fewest counts to execute the video generation task; if the counts are the same, the engine is selected according to the preset service parameters, which include queuing time and billing tier.

[0050] It is understandable that the expected hit rate only affects the order in which engines are selected, not whether an engine is retained. The absence of expected items will not cause the wrong exclusion of available engines. The specific form of the selection rules is not limited, and any rule that can achieve the best selection among the unexcluded engines can be adopted.

[0051] Specifically, the expression replacement module includes: The replacement query submodule is used to identify items in the required content list that are the same as those in the missing items list as items to be replaced, and to determine the replacement rule corresponding to the item to be replaced based on a preset expression replacement lookup table; the expression replacement lookup table stores the correspondence between item types and replacement rules; The representation generation submodule is used to generate equivalent representation entries corresponding to the entries to be replaced according to the replacement rules, and write the equivalent representation entries back to the required content list; The equivalent representation entry inherits the category code of the entry to be replaced.

[0052] Specifically, the expression substitution lookup table stores the correspondence between entry types and substitution rules. The substitution rules are used to change the expression of an entry while keeping its semantics unchanged.

[0053] Please see Figure 4 The replacement query submodule identifies items in the essential content list that are identical to those in the missing items list as items to be replaced. Based on the item type of the item to be replaced, it queries the corresponding replacement rules. The description generation submodule generates equivalent description items according to the replacement rules. The equivalent description items inherit the category code of the item to be replaced and are written back to the essential content list.

[0054] In this embodiment, the replacement rules include: replacing attribute description entries with scene composition description entries containing the attribute; replacing quantity description entries with description entries that present quantity relationships with reference objects; and replacing temporal action description entries with description entries presented as key shot sequences. After the equivalent description entries are written back, the encoding extraction submodule re-extracts the category codes of each entry, and the matching judgment submodule re-executes the comparison with the missing item list of each candidate video generation engine.

[0055] It is understandable that the specific content of the replacement rules is not limited. They can be pre-configured based on the expression response characteristics of each video generation engine, or supplemented by historical records of successful replacements. Since equivalent expression items inherit category codes, the re-comparison is still performed according to the code consistency, and no additional semantic operations need to be introduced.

[0056] Specifically, the expression replacement module prevents video generation tasks from entering the generation scheduling process in the following ways: Record the number of replacements performed on the essential content list; If the number of replacements reaches the preset replacement limit, and there are still no candidate video generation engines that have not been excluded after replacement, then the video generation task is prohibited from entering the generation schedule, an infeasibility receipt is generated, and the infeasibility receipt is returned to the autonomous intelligent agent. The infeasibility receipt contains information about the same items in the required list as those in the missing items list.

[0057] Specifically, the scheduling generation module includes: The generation execution submodule is used to schedule the candidate video generation engines that have not been excluded to perform video generation tasks in order to generate the target video; The finished product inspection submodule is used to perform item inspection on the target video and filter out items from the list of essential content that do not appear in the target video. The segment regeneration submodule is used to determine the time period corresponding to the selected items in the video generation task, and to schedule the candidate video generation engine to regenerate the corresponding video segment for the time period.

[0058] In this embodiment, the segment regeneration submodule maps the selected entries to the corresponding scenes in the video generation task, determines the time period corresponding to the scene in the target video, and only schedules the candidate video generation engine to regenerate the corresponding video segments for the scene, while the video segments corresponding to the other scenes remain unchanged; when regenerating the video segments, the original layout style of the scene is maintained; the regenerated video segments are spliced ​​with the original video segments of the other scenes according to the same timeline to obtain the updated target video.

[0059] It is understandable that replacing the entire video with segment-level regeneration reduces the recalculation cost when an item fails the detection from the entire video duration to the duration of individual scenes. Since the target video is obtained by deterministic rendering from the video compositing code, the video segments regenerated for a single scene can be accurately spliced ​​with the video segments of other scenes along the timeline, which will not be elaborated here.

[0060] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A video generation service system for large language models supporting autonomous intelligent agents, characterized in that, include: The capability detection module is used to schedule the video generation engine to execute a preset set of detection tasks in response to the fulfillment of preset detection trigger conditions, so as to determine the generation results of each detection. The image registration module is used to perform item detection on the generated results to filter out items that do not appear in the generated results from the detection task set, and register the filtered items as a list of missing items in the video generation engine. The task receiving module is used to receive video generation tasks submitted by autonomous intelligent agents and determine a list of essential content based on the video generation tasks. The pre-interception module is used to compare the items in the essential content list with the omission list of each candidate video generation engine before the video generation task enters the generation scheduling. If the omission list of a candidate video generation engine contains an item in the essential content list, the candidate video generation engine is excluded. The expression replacement module is used in response to the exclusion of all candidate video generation engines. Based on the preset expression replacement comparison table, it replaces the items in the essential content list that are the same as the missing items list with the corresponding equivalent expression items, and re-executes the comparison based on the replaced essential content list. If there are still no candidate video generation engines that have not been excluded after replacement, the video generation task is prohibited from entering the generation schedule. The generation scheduling module is used to schedule the candidate video generation engines that have not been excluded to perform video generation tasks in order to generate the target video.

2. The large language model video generation service system supporting autonomous intelligent agents according to claim 1, characterized in that, The process by which the capability detection module determines whether the preset detection triggering conditions are met includes: Monitor the access status and engine identifier of the video generation engine; If the access of the video generation engine is detected or the engine identifier of the video generation engine is changed, it is determined that the preset detection triggering conditions are met.

3. The large language model video generation service system supporting autonomous intelligent agents according to claim 1, characterized in that, The portrait registration module includes: An entry detection submodule is used to perform entry detection on the generated result to determine the occurrence of each entry in the detection task set in the generated result; The missing item filtering submodule is used to filter items that do not appear in the generated results based on their occurrence. The list registration submodule is used to register the item as a missing item list of the video generation engine according to the category code when the number of consecutive absences of an item reaches a preset registration number. The omission list is stored in association with the engine identifier of the video generation engine.

4. The large language model video generation service system supporting autonomous intelligent agents according to claim 1, characterized in that, The task receiving module includes: The task parsing submodule is used to perform content parsing on the video generation task to determine the content items that the video generation task requires to be presented; The coding and labeling submodule is used to assign category codes to content items based on the category coding system and organize the content items with assigned category codes into a list of essential content. The necessity labeling submodule is used to determine the necessity of each item in the essential content list; The necessity indicator includes a necessity level and a desired level.

5. The large language model video generation service system supporting autonomous intelligent agents according to claim 1, characterized in that, The pre-interception module includes: The encoding extraction submodule is used to extract the category codes of each item in the essential content list and to extract the category codes of each item in the missing item list of each candidate video generation engine. The matching and determination submodule is used to determine that items with the same category code are the same items, and based on the determination result of the same items, it determines whether the omission list contains items from the required content list. The items in the essential content list and the missing items list are determined based on the same category coding system.

6. The large language model video generation service system supporting autonomous intelligent agents according to claim 5, characterized in that, The items in the required list are marked with a necessity indicator; The process by which the pre-interception module excludes candidate video generation engines includes: If the omission list of candidate video generation engines contains items marked as necessary, then the candidate video generation engine is excluded. If the list of missing items in the candidate video generation engine contains items marked as "desired" in terms of necessity, then the desired level hit rate of the candidate video generation engine is recorded, and the candidate video generation engine is retained.

7. The large language model video generation service system supporting autonomous intelligent agents according to claim 6, characterized in that, The process by which the generation scheduling module schedules candidate video generation engines that have not been excluded includes: Obtain the expected hit rate of candidate video generation engines that have not been excluded; The candidate video generation engine with the fewest expected hits is selected to perform the video generation task.

8. The large language model video generation service system supporting autonomous intelligent agents according to claim 1, characterized in that, The expression replacement module includes: The replacement query submodule is used to identify items in the required content list that are the same as those in the missing items list as items to be replaced, and to determine the replacement rule corresponding to the item to be replaced based on a preset expression replacement lookup table; the expression replacement lookup table stores the correspondence between item types and replacement rules; The representation generation submodule is used to generate equivalent representation entries corresponding to the entries to be replaced according to the replacement rules, and write the equivalent representation entries back to the required content list; The equivalent representation entry inherits the category code of the entry to be replaced.

9. The large language model video generation service system supporting autonomous intelligent agents according to claim 1, characterized in that, The process by which the expression replacement module prevents video generation tasks from entering the generation schedule includes: Record the number of replacements performed on the essential content list; If the number of replacements reaches the preset replacement limit, and there are still no candidate video generation engines that have not been excluded after replacement, then the video generation task is prohibited from entering the generation schedule, an infeasibility receipt is generated, and the infeasibility receipt is returned to the autonomous intelligent agent. The infeasibility receipt contains information about the same items in the required list as those in the missing items list.

10. The large language model video generation service system supporting autonomous intelligent agents according to claim 1, characterized in that, The generation scheduling module includes: The generation execution submodule is used to schedule the candidate video generation engines that have not been excluded to perform video generation tasks in order to generate the target video; The finished product inspection submodule is used to perform item inspection on the target video and filter out items from the list of essential content that do not appear in the target video. The segment regeneration submodule is used to determine the time period corresponding to the selected items in the video generation task, and to schedule the candidate video generation engine to regenerate the corresponding video segment for the time period.

Citation Information

Patent Citations

  • A video generation method, system, device and medium based on a large language model

    CN120378710B