Video generation methods, devices, media, and products based on agent-based iterative optimization

By constructing a directed graph to record the multi-agent generation process and transforming user evaluations into structured solutions, detecting and reconstructing the orchestration structure, and diagnosing problematic agents, the problem of user evaluations not being adopted in existing technologies is solved, and standardized evaluation and efficient optimization of film generation are achieved.

CN122802757BActive Publication Date: 2026-11-10BEIJING ELECTRONIC DIGITAL INTELLIGENCE TECHNOLOGY CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611265613.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-10
Estimated Expiration
2046-08-20

AI Technical Summary

Technical Problem

In existing AIGC video automated production systems based on large language models and multimodal models, user feedback is not fully adopted, resulting in generated videos that fail to meet actual user needs. Furthermore, the multi-agent iterative self-optimization method relies on fixed standards and ignores users' subjective perceptions.

Method used

By constructing a directed graph to record the multi-agent generation process, user evaluations are obtained and transformed into a structured evaluation scheme. The program structure is detected, the program logic is reconstructed, problematic agents are diagnosed and optimized, until a film that meets user expectations is generated.

Benefits of technology

It achieves standardized evaluation of the film generation process, automatically identifies mismatches in the arrangement logic or insufficient capabilities of the intelligent agent, improves generation efficiency and quality, reduces the frequency of human intervention, and continuously optimizes film quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122802757B_ABST
    Figure CN122802757B_ABST
Patent Text Reader

Abstract

This disclosure presents a video generation method, apparatus, medium, and product based on agent-based iterative optimization. The method includes: selecting multiple specialized agents to generate a video based on user-inputted video generation prompts, and recording the generation process trajectory using a directed graph; converting user feedback on the video into an evaluation scheme, and when the video is determined to be substandard, detecting the arrangement structure between the specialized agents based on the current directed graph and historical directed graphs in the process pattern library, and reconstructing the structure if the detection fails; after passing the detection, performing specialized agent diagnosis based on the evaluation scheme, and optimizing problematic agents; and regenerating the video based on the optimized specialized agents and the reconstructed arrangement structure until convergence is achieved, and updating the process pattern library. This method allows video optimization to be directly adjusted based on real user feedback, no longer limited by fixed standards, making the regenerated video more in line with user expectations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to the field of AIGC video generation technology, and in particular to a video generation method, apparatus, medium and product based on intelligent agent iterative optimization. Background Technology

[0002] AIGC video automated production systems based on large language models and multimodal models typically consist of multiple intelligent agents working together, including script generation, visual generation, audio generation, and editing, to achieve end-to-end automated video production. Since the output quality of preceding stages directly affects subsequent stages during the sequential execution of these agents, and video generation involves multiple evaluation dimensions such as narrative logic, visual presentation, and audio matching, substandard generation quality in any stage can lead to a poor final product. Therefore, it is necessary to investigate each agent to pinpoint the source of the problem and implement targeted optimization.

[0003] Currently, a multi-agent iterative self-optimization method is commonly used for screening and tuning. This method uses a specialized evaluation agent to evaluate the generated video in multiple dimensions (such as picture quality, audio synchronization, and content coherence). Then, the inference agent converts the evaluation results into optimization instructions to drive the next round of generation iterations until the preset maximum number of iterations is reached or the scores of the multi-dimensional evaluations reach a preset threshold.

[0004] The conclusion that the final product is not good usually comes from user reviews. However, the current technology's troubleshooting and optimization process relies entirely on preset internal evaluation standards, ignoring the evaluation content input by users. The subjective perception of poor performance by users is not adopted in the troubleshooting and optimization process, resulting in the regenerated film still failing to meet the actual needs of users. Summary of the Invention

[0005] In view of this, the present disclosure provides a video generation method, apparatus, medium and product based on intelligent agent iterative optimization, which enables video optimization to be directly adjusted according to users' real evaluations, no longer limited by fixed standards, and makes the regenerated video more in line with user expectations.

[0006] In a first aspect, embodiments of this disclosure provide a video generation method based on agent-based iterative optimization, employing the following technical solution: Obtain the video generation prompt words input by the user, select multiple specialized intelligent agents to generate videos based on the video generation prompt words, and use a directed graph to record the trajectory of the multiple specialized intelligent agents in the video generation process; Obtain user reviews of the film and transform those reviews into a structured evaluation scheme; When it is determined that the film does not meet the standard based on the evaluation scheme, the arrangement structure between special intelligent agents is detected based on the current directed graph and the historical directed graph in the process pattern library; If the detection fails, the orchestration structure between the specialized intelligent agents is reconstructed; After the test is passed, a special intelligent agent diagnosis is performed based on the evaluation scheme to determine the problem type and the problem intelligent agent, and the problem intelligent agent is optimized based on the problem type; Based on the optimized specialized intelligent agent and the reconstructed arrangement structure, the film is regenerated until the convergence condition is met. Based on the optimized specialized intelligent agents and the reconstructed orchestration structure, the current directed graph is corrected, and the corrected directed graph is stored in the process pattern library.

[0007] Optionally, the step of transforming the manual evaluation into a structured evaluation scheme includes: Construct an evaluation dimension table, which includes multiple special dimensions and evaluation indicators, evaluation indicator weights, evaluation indicator thresholds, and evaluation indicator collection methods for each special dimension; The manual evaluation is broken down into multiple sentences. From the evaluation dimension table, the specific dimension to which each sentence belongs is identified, and the problem description is extracted. According to the evaluation index collection method of each specific dimension, the problem description of each sentence is analyzed to obtain the corresponding evaluation index values; The evaluation index values, weights, thresholds, and data collection methods for each sentence are integrated into a structured evaluation scheme.

[0008] Optionally, the detection of the orchestration structure between specialized intelligent agents based on the current directed graph and historical directed graphs in the process pattern library includes: Obtain the current movie type, and filter the historical directed graph of the same movie type from the process pattern library; Obtain the structural similarity between the current directed graph and the filtered historical directed graphs; When the structural similarity is greater than a preset similarity threshold, the detection is deemed to have passed; when the structural similarity is not greater than the preset similarity threshold, the detection is deemed to have failed.

[0009] Optionally, the reconfiguration of the orchestration structure between the specialized intelligent agents includes: Select the historical directed graph with the highest similarity to the current directed graph structure from the process pattern library as the reference directed graph; The current directed graph is compared with the reference directed graph to identify differences. The differences include at least one of the following: missing or redundant agent nodes, differences in node execution order, inconsistencies in serial-parallel relationships, and mismatched branch conditions. Based on the differences, a target reconstruction operation is determined from multiple preset reconstruction operations. The reconstruction operation includes at least one of adding or deleting agents, adjusting the execution order, switching between serial and parallel modes, and modifying branch conditions. The target reconstruction operation is performed on the orchestration structure between the specialized intelligent agents, and the legality of the reconstructed directed graph is verified. Once the validity check passes, the reconstructed orchestration structure is re-checked, and once the check passes, the reconstructed orchestration structure is output.

[0010] Optionally, the step of performing specialized agent diagnosis based on the evaluation scheme to determine the problem type and the problem agent includes: The difference between the value of each evaluation indicator in the evaluation scheme and the corresponding evaluation indicator threshold is obtained, and the difference is weighted by the corresponding evaluation indicator weight to obtain the loss value of each special dimension. Obtain the actual output and baseline output of each specialized intelligent agent, wherein the baseline output is the historical normal output data of the same type of intelligent agent as the specialized intelligent agent extracted from the reference directed graph; Iterate through each specialized agent, replace the actual output of the currently traversed specialized agent with the baseline output, keep the links of other specialized agents unchanged, and re-obtain the loss value of the specialized dimension to which the specialized agent belongs. Use the change in the loss value before and after the replacement as the loss contribution of the specialized agent. When the loss contribution of the specialized intelligent agent is greater than the preset contribution threshold, the problem type is determined based on the type of the specialized intelligent agent, and the specialized intelligent agent is identified as the problem intelligent agent.

[0011] Optionally, the specific dimension includes an emotion dimension; the step of performing specific agent diagnosis based on the evaluation scheme to determine the problem type and the problem agent further includes: When the loss value of the emotion dimension is greater than the preset loss threshold, all special agents that affect the generation of the emotion curve of the film are identified as problem agents, and the problem type is determined to be a cross-agent problem.

[0012] Optionally, the optimization of the problem agent based on the problem type includes: When the problem type is a cross-agent problem, obtain the sentiment curve of the film; Detect the emotional plateau interval on the emotional curve, and record the start and end times and emotional values ​​of the emotional plateau interval; Based on the start and end times of the emotion plain interval and the difference in emotion values ​​before and after the emotion plain interval, the linkage instructions for each problem agent are matched from the instruction template library. Each linkage command is sent to the corresponding problem agent for optimization.

[0013] Secondly, this disclosure also provides a video generation system based on intelligent agent iterative optimization, employing the following technical solution: The recording module is used to obtain the video generation prompt words input by the user, select multiple specialized intelligent agents to generate videos based on the video generation prompt words, and use a directed graph to record the trajectory of the multiple specialized intelligent agents in the video generation process. The conversion module is used to obtain users' human evaluations of the film and convert the human evaluations into a structured evaluation scheme. The detection module is used to detect the arrangement structure between special intelligent agents based on the current directed graph and the historical directed graph in the process pattern library when it is determined that the film does not meet the standard based on the evaluation scheme. The reconstruction module is used to reconstruct the orchestration structure between the specialized intelligent agents if the detection fails. The optimization module is used to perform specific intelligent agent diagnosis based on the evaluation scheme after the detection is passed, determine the problem type and the problem intelligent agent, and optimize the problem intelligent agent based on the problem type; The generation module is used to regenerate the film based on the optimized specialized intelligent agent and the reconstructed orchestration structure until the convergence condition is met. The correction module is used to correct the current directed graph based on the optimized special agent and the reconstructed orchestration structure, and store the corrected directed graph in the process pattern library.

[0014] Thirdly, this disclosure also provides a computer device, which adopts the following technical solution: The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform any of the above-described agent-based iterative optimization-based video generation methods.

[0015] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing computer instructions for causing a computer to execute any of the above-described agent-based iterative optimization-based video generation methods.

[0016] Fifthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0017] The video generation method based on agent-based iterative optimization provided in this disclosure records the entire process of multi-agent collaborative video generation by constructing a directed graph, providing an analyzable structured basis for subsequent optimization. It transforms subjective user evaluation into an objective and quantifiable structured evaluation scheme, shifting video quality judgment from experience-based reliance to standardized assessment. This allows for the automatic identification of whether video defects stem from mismatched arrangement logic between agents or insufficient capabilities of specific agents. Furthermore, by comparing the current directed graph with a historical success case library, the arrangement structure is dynamically detected and reconstructed to ensure the rationality of agent collaboration. After arrangement is approved, the problematic agent and its specific type are accurately diagnosed, enabling targeted optimization and avoiding inefficient adjustments to all agents. Through iterative generation, verification, and convergence, not only are videos that meet user expectations automatically produced, but the optimized directed graph is also fed back to the process pattern library, achieving knowledge accumulation and reuse. This significantly reduces the frequency of manual intervention and debugging costs while continuously improving generation efficiency and video quality.

[0018] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the video generation method based on agent-based iterative optimization provided in this embodiment of the disclosure; Figure 2 A flowchart illustrating the evaluation scheme acquisition method provided in this embodiment of the disclosure; Figure 3 A flowchart illustrating the method for detecting the arrangement structure between specialized intelligent agents provided in this embodiment of the disclosure; Figure 4 A flowchart illustrating the specific intelligent agent orchestration structure reconstruction method provided in this embodiment of the disclosure; Figure 5 This is a schematic flowchart of a specific intelligent agent diagnostic method provided in an embodiment of this disclosure; Figure 6 A flowchart illustrating the agent optimization method for cross-agent problems provided in this embodiment of the disclosure; Figure 7 A schematic diagram of a video generation system based on agent-based iterative optimization provided in this embodiment of the disclosure; Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0021] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0022] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0023] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0024] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0025] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0026] Reference Figure 1 This disclosure provides a video generation method based on agent-based iterative optimization, comprising the following steps: S1: Obtain the video generation prompt words input by the user, select multiple specialized intelligent agents to generate videos based on the video generation prompt words, and use a directed graph to record the trajectory of the video generation process of multiple specialized intelligent agents. S2: Obtain user reviews of the film and transform these reviews into a structured evaluation scheme; S3: When a film is determined to be substandard based on the evaluation scheme, the arrangement structure between specialized intelligent agents is detected based on the current directed graph and the historical directed graph in the process pattern library; S4: If the detection fails, the orchestration structure between the special intelligent agents will be reconstructed; S5: After the test is passed, perform a special intelligent agent diagnosis based on the evaluation plan to determine the problem type and the problem intelligent agent, and optimize the problem intelligent agent based on the problem type; S6: Based on the optimized specialized intelligent agent and the reconstructed arrangement structure, regenerate the film until the convergence condition is met; S7: Based on the optimized specialized intelligent agent and the reconstructed orchestration structure, the current directed graph is corrected, and the corrected directed graph is stored in the process pattern library.

[0027] This disclosed method for video generation based on agent-based iterative optimization records the entire process of multi-agent collaborative video generation by constructing a directed graph, providing an analyzable structured basis for subsequent optimization. It transforms subjective user evaluation into an objective and quantifiable structured evaluation scheme, shifting video quality judgment from experience-based reliance to standardized assessment. This allows for the automatic identification of whether video defects stem from mismatched arrangement logic between agents or insufficient capabilities of specific agents. Furthermore, by comparing the current directed graph with a historical success case library, the arrangement structure is dynamically detected and reconstructed to ensure the rationality of agent collaboration. After successful arrangement, it accurately diagnoses problematic agents and their specific types, enabling targeted optimization and avoiding inefficient adjustments to all agents. Through iterative generation, verification, and convergence, it not only automatically produces videos that meet user expectations but also feeds the optimized directed graph back to the process pattern library, achieving knowledge accumulation and reuse. This significantly reduces the frequency of manual intervention and debugging costs while continuously improving generation efficiency and video quality.

[0028] In S1, the film generation agent consists of multiple agents, including script generation, visual generation, audio generation, and editing, arranged in a dynamic process. These agents interact through a standardized interface protocol, which defines the input parameter structure, output product structure, metadata fields, and exception status codes. The input parameters for each agent are divided into two categories: required parameters and optional parameters. Required parameters are the prerequisite data upon which the agent's execution depends. If optional parameters are missing, predefined default values ​​in the protocol are automatically used.

[0029] The method parses user-inputted video generation prompts and selects the appropriate agent to generate the video. When subsequent adjustments to the arrangement of these agents are needed, compatibility checks are performed by comparing the input and output types of the agents before and after. Each agent's mandatory inputs must have corresponding outputs from its predecessor. If an upstream agent is removed, causing a downstream agent's input to be unavailable, the input will be filled with a default value if it's optional, and if it's mandatory, a validity check will be triggered, indicating the need to add a replacement agent or adjust the data link. This method allows for flexible agent arrangement to adapt to different user themes, without being limited by a fixed calling order due to interface coupling. This ensures that the video generation prompts based on user input can flexibly select suitable agent combinations and guarantee their operability.

[0030] During the video generation process, the generation trajectory is recorded using a directed graph structure. The data is persistently stored in the form of an immutable log and is fully backtracked and reconstructed according to the video instance identifier. This directed graph consists of three core components: a node set, an edge set, and a global context. The node set records the agent type identifier, version number, input parameter snapshot, output product summary, execution time, resource consumption, and configuration parameters for each specialized agent call instance. The edge set records the data dependencies between specialized agents and the data flow path of intermediate products. The global context records the user intent tags, style constraint parameters, theme type identifiers, and mid-process intervention instructions during the process execution.

[0031] For example, the script generation agent records the input theme, style constraints, and duration requirements, and outputs the script, storyboard list, emotion tags, and character list; the visual generation agent records the input storyboard description, style keywords, and aspect ratio, and outputs image descriptive words, model parameters, seed values, and image color / brightness statistical features; the audio generation agent records the input emotion tags, duration, and rhythm requirements, and outputs music emotion tags, BPM (Beats Per Minute), instrument type, and volume envelope; the editing agent records the input material list and timeline, and outputs shot duration, transition sequence, and cut point alignment markers with music beats. All records are linked together through a directed graph to form a complete data chain, achieving a structured record of the complete process trajectory of multiple specialized agents generating films.

[0032] In S2, after the initial video is generated, users will provide subjective evaluations, including ratings, textual feedback, and modification tags. (See reference...) Figure 2 The flowchart illustrating the method for obtaining evaluation schemes shows that "converting manual evaluations into structured evaluation schemes" includes the following steps: S21: Construct an evaluation dimension table, which includes multiple special dimensions and evaluation indicators for each special dimension, evaluation indicator weights, evaluation indicator thresholds, and evaluation indicator collection methods; S22: Break down the manual evaluation into multiple sentences, identify the specific dimension to which each sentence belongs from the evaluation dimension table, and extract the problem description; S23: According to the evaluation index collection method of each sentence's specific dimension, analyze the problem description of each sentence and obtain the corresponding evaluation index values; S24: Integrate the evaluation index values, evaluation index weights, evaluation index thresholds, and evaluation index collection methods for each sentence into a structured evaluation scheme.

[0033] In S21, an evaluation dimension table covering all dimensions of film generation is pre-constructed. This table includes five specific dimensions: script, visuals, music, rhythm, and emotion. Each specific dimension contains several evaluation indicators. Among them, the script dimension corresponds to the number of conflict points, the number of turning points, and the dialogue density ratio per scene; the visual dimension corresponds to the overlap rate of style keywords, the consistency of character seeds, and the brightness / color distribution of the screen; the music dimension corresponds to the matching rate of audio-visual emotion tags, the alignment rate of BPM with editing points, and the standard deviation of volume; the rhythm dimension corresponds to the number of shots in the first N seconds, the average shot duration, and the distribution of transition types; and the emotion dimension corresponds to the sequential emotion value of the entire film and the duration of the emotion plateau.

[0034] Each evaluation indicator includes three attributes: weight, threshold, and data collection method. The weight is the proportion of the indicator in the score of its respective dimension. The sum of the weights of all indicators in the same dimension is 1. The weight is initially set by the experience of film and television experts and is continuously calibrated based on the qualified cases in the process pattern library. The threshold is the critical value for determining whether the indicator meets the standard. It is determined based on the statistical distribution of the indicator in historical qualified cases (such as the quartile values ​​above). The data collection method specifies how to obtain the specific value of the indicator based on each question description in the manual evaluation. It is divided into two categories: direct statistics from trajectory logs and model inference.

[0035] Direct statistics from trajectory logs refer to extracting data from the structured log records of each specialized intelligent agent node. For example, extracting the number of shots, shot duration, and transition type from the editing node log, and the number of dialogue words from the script generation node log, thereby quantifying the problem description (such as too few shots, abrupt transitions, and too little dialogue) into the corresponding indicator values. Model inference refers to using pre-trained models to analyze the intermediate products and final output generated by each specialized intelligent agent to infer the indicator values. For example, using pre-trained models to comprehensively analyze the emotion tag sequence, screen color brightness, and music features to generate temporal emotion values ​​and detect emotion plateaus, and using scene-by-scene script analysis to identify conflict points, thereby quantifying the problem description (such as too flat emotions and lack of conflict) into the corresponding indicator values.

[0036] In S22, a large language model is used to perform semantic parsing on unstructured, manually input evaluation text. The entire evaluation text input by the user is broken down into multiple independent sentences according to semantic units, so that each sentence carries a single evaluation intent. Intent extraction and dimension classification are performed on each segmented sentence to identify the specific dimension to which the sentence belongs and its corresponding specific problem description.

[0037] For example, the evaluation statement "too drawn out in the middle, lacking conflict" is analyzed as "insufficient plot density" and categorized under the script dimension; "the style of the third image is too realistic and inconsistent with other scenes" is analyzed as "stylistic consistency deviation" and categorized under the visual dimension; "the ending music is too cheerful and clashes with the farewell atmosphere" is analyzed as "mismatch of emotional labels" and categorized under the music dimension; "the beginning is slow-paced and fails to grab attention" is analyzed as "insufficient density of opening shots" and categorized under the pacing dimension; and "the overall experience is too bland, making me want to quit halfway through" is analyzed as "flat emotional curve" and categorized under the emotional dimension. This analysis process maps the natural language description of the human evaluation to the predefined specific dimensions and question types in the evaluation dimension table.

[0038] In S23, for each question description, the specific values ​​of the evaluation indicators are obtained according to the evaluation indicator collection method of its respective special dimension.

[0039] For example, the evaluation metrics are collected directly from the trajectory logs. To describe the problem of too few shots, the number of cuts within the first N seconds is counted from the editing node logs to obtain the shot count index for the first N seconds. To describe the problem of sluggish shots, the average duration of all shots is calculated from the editing node logs to obtain the average shot duration index. To describe the problem of abrupt transitions, the frequency and proportion of various transition effects are counted from the editing node logs to obtain the transition type distribution index. To describe the problem of too little dialogue, the proportion of dialogue words to total words is counted from the script generation node logs to obtain the dialogue density ratio index.

[0040] For example, the evaluation metrics are collected through model inference. For the problem of emotional flatness, the emotional label sequence output by the script generation agent, the color and brightness data of the visual generation agent, and the music emotional labels and volume envelope output by the audio generation agent are input into a pre-trained model. After fusion analysis, a full-film temporal emotion curve is generated, and the duration of the emotional plateau is detected to obtain the full-film temporal emotion value and the duration of the emotional plateau. For the problem of lacking conflict, the script output by the script generation agent is input into the pre-trained model scene by scene. The number of emotional reversals and target conflicts in each scene, as well as the number of nodes with significant changes in plot direction, are counted to obtain the number of conflict points and the number of turning points in each scene. For the problem of inconsistent style, the emotional labels of each scene from the visual generation agent are input into the pre-trained model. Input style keywords into the pre-trained model and compare the overlap ratio of style keywords when different scenes are generated to obtain the style keyword overlap rate index. For the problem of characters not looking like the same person, the consistency of seed values ​​and reference image labels in each scene is judged by model comparison to obtain the character seed consistency index. For the problem of music emotion not matching, the expected emotion label of the scene and the music emotion label output by the audio generation agent are input into the pre-trained model, and the label consistency ratio is calculated to obtain the audio-visual emotion label matching rate index. For the problem of editing and music being out of sync, the BPM and beat sequence output by the audio generation agent and the editing timeline output by the editing agent are input into the pre-trained model, and the alignment degree is calculated to obtain the BPM and editing point alignment rate index.

[0041] In S24, the acquired evaluation index values, along with the weights, thresholds, and collection methods of the corresponding indicators preset in the evaluation dimension table, are grouped and integrated into an evaluation scheme according to the specific dimensions.

[0042] In S3, for the currently generated video, the difference between the value of each evaluation indicator in the evaluation scheme and the corresponding evaluation indicator threshold is calculated. The difference is weighted by the corresponding evaluation indicator weight to obtain the loss value of that specific dimension. Then, the loss values ​​of each specific dimension are summarized to obtain the comprehensive evaluation score. The comprehensive evaluation score is compared with the preset compliance threshold, and it is checked whether each specific dimension meets the threshold requirements. If the comprehensive evaluation score reaches the preset compliance threshold and no single item in each dimension fails to meet the standard, the video is judged to be qualified, and the current directed graph of the video is stored in the process pattern library as a new reference case; otherwise, the video is judged to be unqualified, and the agent tuning task is automatically triggered.

[0043] Reference Figure 3 The flowchart illustrating the method for detecting the arrangement structure between specialized intelligent agents, which "detects the arrangement structure between specialized intelligent agents based on the current directed graph and historical directed graphs in the process pattern library," includes the following steps: S31: Obtain the current movie type and filter the historical directed graph of the same movie type from the process pattern library; S32: Obtain the structural similarity between the current directed graph and the filtered historical directed graphs; S33: When the structural similarity is greater than the preset similarity threshold, the detection is deemed to have passed; when the structural similarity is not greater than the preset similarity threshold, the detection is deemed to have failed.

[0044] The process pattern library pre-stores historical optimization cases. Each record contains the directed graph structure of the case (including agent sequence, branch logic, and serial-parallel relationship), film type label, loss value of each special dimension, and comprehensive evaluation score.

[0045] Based on the user-inputted video prompts, the system identifies the type of the currently generated video and filters historical directed graphs corresponding to videos of the same type from the workflow pattern library. The system then performs a topological comparison between the current video's directed graph and each of the filtered historical directed graphs, calculating structural similarity using node type matching rate and edge relationship consistency as metrics.

[0046] The maximum structural similarity value among all comparison results is obtained. When the maximum structural similarity is greater than the preset similarity threshold, the detection is deemed to have passed, indicating that the current process structure is close enough to the historical successful pattern, and there is no need to reconstruct the orchestration structure between the special agents. The process can directly enter the agent diagnosis process. When the maximum structural similarity is not greater than the preset similarity threshold, the detection is deemed to have failed, indicating that the current process structure is significantly different from all historical successful patterns, and the orchestration structure reconstruction process needs to be triggered.

[0047] In addition, if the current structural similarity is greater than the preset similarity threshold, but the manual evaluation involves negative descriptions at the level of the special intelligent agent orchestration structure (such as "unreasonable module matching" or "lack of structure fluctuations"), it means that the special intelligent agent orchestration structure itself is reasonable, but the execution quality of the intelligent agent is insufficient, and the special intelligent agent orchestration structure is still allowed to pass the detection.

[0048] In S4, refer to Figure 4 The flowchart illustrating the method for reconstructing the orchestration structure of specialized intelligent agents shows that "reconstructing the orchestration structure between specialized intelligent agents" includes the following steps: S41: Select the historical directed graph with the highest similarity to the current directed graph structure from the process pattern library as the reference directed graph; S42: Compare the current directed graph with the reference directed graph to identify topological differences. The differences include at least one of the following: missing agent nodes, differences in node execution order, inconsistencies in serial-parallel relationships, and mismatched branch conditions. S43: Based on the difference item, determine the target reconstruction operation from multiple preset reconstruction operations. The reconstruction operation includes at least one of adding or deleting agents, adjusting the execution order, switching between serial and parallel modes, and modifying branch conditions. S44: Perform the target reconstruction operation on the current directed graph and perform a validity check on the reconstructed directed graph; S45: After the validity check passes, the reconstructed orchestration structure is re-checked. After the check passes, the reconstructed orchestration structure is output.

[0049] In S42, the current directed graph and the reference directed graph are compared node-by-node and edge-by-edge in terms of topology. Specifically, the node sets are compared to identify agent node types that exist in the reference directed graph but are missing in the current directed graph, and these are marked as missing agent nodes. Agent node types that do not exist in the reference directed graph but exist in the current directed graph are marked as redundant agent nodes. The execution order between nodes is compared to identify whether there are differences in the execution order of the same type of agent nodes in the two graphs, and these are marked as differences in node execution order. The parallel dependencies between nodes are compared to identify data dependencies that are executed in parallel in the reference directed graph but sequentially in the current directed graph (or sequentially in the reference directed graph but in parallel in the current directed graph), and these are marked as inconsistent serial-parallel relationships. The branch decision logic of conditional branch nodes is compared with the execution path after the branch, and branch conditions or branch targets that exist in the reference directed graph but are missing in the current directed graph are marked as branch condition mismatches. The above differences are categorized and summarized by type, forming the input basis for subsequent reconstruction operations.

[0050] In S43, the system analyzes whether there are process problem descriptions at the level of specific agent execution order in the human evaluation. If so, it combines the identified discrepancies with the process problem descriptions to determine the target reconstruction operation. When the discrepancy includes missing or redundant agent nodes and the human evaluation explicitly mentions the absence or redundancy of a certain step, it determines the reconstruction operation of adding or deleting agents to fill in the missing nodes. When the discrepancy includes differences in the execution order of nodes and the human evaluation mentions incorrect structural fluctuations, it determines the reconstruction operation of adjusting the execution order to rearrange the execution sequence of each node. When the discrepancy includes inconsistent serial-parallel relationships and the human evaluation mentions loose rhythm, it determines the reconstruction operation of switching the serial-parallel mode to optimize data dependencies. When the discrepancy includes mismatched branch conditions and the human evaluation mentions insufficient coverage, it determines the reconstruction operation of modifying branch conditions to expand scene coverage.

[0051] If the manual evaluation does not describe the process problem, the corresponding reconstruction operation is determined directly based on the type of discrepancy. For example, missing agent nodes correspond to adding or deleting agents; differences in node execution order correspond to adjusting the execution order; inconsistent serial-parallel relationships correspond to switching to serial-parallel mode; and mismatched branch conditions correspond to modifying the branch conditions. When multiple discrepancies exist simultaneously, the corresponding reconstruction operations are determined sequentially according to a preset priority (execution order takes precedence over agent addition / removal, agent addition / removal takes precedence over serial-parallel switching, and serial-parallel switching takes precedence over branch condition modification), forming a reconstruction operation sequence.

[0052] In S44, the orchestration structure between specialized agents is modified according to the determined target reconstruction operation (or reconstruction operation sequence). For example, based on the agent addition / deletion operation, new specialized agent nodes are added or existing specialized agent nodes are removed from the orchestration structure, and the corresponding nodes and their associated edges are added or deleted in the directed graph simultaneously. Based on the execution order adjustment operation, the execution order of each specialized agent in the orchestration structure is rearranged, and the execution order of nodes in the directed graph is adjusted simultaneously. Based on the serial-parallel mode switching operation, the dependency relationship type between each specialized agent in the orchestration structure is modified, changing serial relationships to parallel relationships or vice versa, and the corresponding edge type in the directed graph is modified simultaneously. Based on the branch condition modification operation, the decision rules and branch execution targets of the conditional branch nodes in the orchestration structure are adjusted, and the corresponding branch conditions and branch paths in the directed graph are modified simultaneously.

[0053] The reconstructed orchestration structure undergoes a validity check, including detecting circular dependencies. This involves traversing the path from any specialized agent node along the dependency relationships to check for paths that return to the original node after passing through several nodes. If a circular dependency exists, the check fails. The system also checks the input and output data types of each specialized agent node. For each mandatory input, it checks if a preceding specialized agent node provides the corresponding output type. If a downstream agent's input is missing due to the removal of an upstream agent, it determines whether the input is mandatory or optional. If optional, it automatically fills in with predefined default values; if mandatory, it triggers a validity check error. Furthermore, it checks whether adding or removing agents after reconstruction causes incompatibility in input and output types between nodes. If the validity check fails, a new target reconstruction operation is selected.

[0054] In S45, if the validity check passes, the orchestration structure is re-checked until the check passes or the preset maximum number of reconstructions is reached, at which point the process is forcibly terminated and a reconstruction failure is reported.

[0055] In S5, refer to Figure 5 The flowchart of the specialized intelligent agent diagnostic method shown includes the following steps: "Performing specialized intelligent agent diagnostics based on the evaluation scheme to determine the problem type and the problematic intelligent agent": S51: Obtain the difference between the value of each evaluation indicator in the evaluation scheme and the corresponding evaluation indicator threshold, and then weight the difference by the corresponding evaluation indicator weight to obtain the loss value of each special dimension; S52: Obtain the actual output and baseline output of each specialized agent. The baseline output is the historical normal output data of agents of the same type as the specialized agent, extracted from the reference directed graph. S53: Traverse each specialized agent, replace the actual output of the currently traversed specialized agent with the baseline output, keep the links and outputs of other specialized agents unchanged, re-acquire the loss value of the specialized dimension to which the specialized agent belongs, and use the change in the loss value before and after the replacement as the loss contribution of the specialized agent. S54: When the loss contribution of a special agent is greater than the preset contribution threshold, the problem type is determined based on the special dimension to which the special agent belongs, and the special agent is identified as the problem agent.

[0056] In S51-S52, the loss values ​​for each specific dimension have already been obtained in step S3 and can be directly called.

[0057] Extract the actual output data of each specialized agent from the current directed graph. The actual output includes the script, storyboard list and emotion tag sequence output by the script generation agent, the scene description words, model parameters, seed values ​​and color and brightness statistical features of each scene output by the visual generation agent, the music emotion tags, BPM, instrument type and volume envelope output by the audio generation agent, and the shot duration sequence, transition sequence and cut point and music beat alignment mark output by the editing agent.

[0058] From the reference directed graph, historical normal output data of the same type of agent is extracted according to the agent type identifier as the benchmark output. The benchmark output represents the normal output level of this type of agent in the process when it produces the standard quality in historical successful cases, and serves as a reference standard for measuring the current agent output deviation in subsequent counterfactual reasoning.

[0059] In S53, counterfactual reasoning is used to perform attribution analysis on each specialized agent. Under the current orchestration structure among specialized agents, one specialized agent is selected as the analysis target, and the actual output of that agent is replaced with the corresponding baseline output, while keeping the link connections of all other specialized agents unchanged.

[0060] After the replacement is complete, the replaced output is used as input to drive the downstream agent to continue execution, regenerating the complete video. The evaluation metrics are recalculated according to the original evaluation scheme's metric collection method, thus obtaining the replacement loss value for the specific dimension belonging to that agent. The recalculated loss value is compared with the original loss value of that dimension before replacement, and the difference is calculated. This difference represents the contribution of the specific agent to the current evaluation loss; a larger difference indicates a greater negative impact of the specific agent on the video's substandard quality. After completing the analysis of one specific agent, the original output of that specific agent is restored, and the next specific agent is selected as the analysis target. The above process is repeated until all specific agents have been traversed, obtaining the loss contribution value for each agent.

[0061] In S54, for specific agents whose loss contribution exceeds the preset contribution threshold, it indicates that their output deviation significantly contributes to the substandard film quality. These agents are then selected as key optimization targets, and the problem type is determined based on the type of the agent. If it is a script generation agent, the problem type is a script problem; if it is a visual generation agent, the problem type is a visual problem; if it is an audio generation agent, the problem type is an audio problem; and if it is an editing agent, the problem type is an editing problem.

[0062] If the problematic agent is a script generation agent, the script optimization unit is invoked for tuning. The script optimization unit triggers corresponding optimization operations based on the compliance status of each indicator in the script dimension of the evaluation scheme, according to the preset diagnostic and optimization mapping rules for script generation agents. If the number of conflict points per scene is insufficient, a conflict density constraint rule is added to the prompts of the script generation agent, requiring each scene to contain a set number of emotional reversals and target conflicts. If the number of plot twists is insufficient, a plot twist requirement is added to the prompts, and an outline verification step is added to check the causal logic coherence of the generated script outline, ensuring that the number of nodes where the plot direction changes significantly meets the requirements. If the dialogue density ratio is insufficient, a range of dialogue words as a percentage of the total number of words is specified in the prompts, constraining the percentage of purely narrative words to achieve a balance between dialogue and narration in the script. Each of the above optimization operations is determined independently and triggered precisely. Related optimization operations are only executed when the corresponding diagnostic item fails to meet the standard. If multiple indicators fail to meet the standard simultaneously, the corresponding multiple optimization operations are executed sequentially.

[0063] If the problem agent is a visual generation agent, the visual optimization unit is invoked for tuning. The visual optimization unit triggers corresponding optimization operations according to the pre-defined visual generation agent diagnosis and optimization mapping rules, based on the compliance status of each visual dimension indicator in the evaluation scheme. If the style keyword overlap rate is not up to standard, indicating insufficient style consistency between images, a global style prefix is ​​added to the input layer of the visual generation agent, and the style parameters are locked to ensure that all images use a unified style constraint. If the character seed consistency is not up to standard, indicating inconsistencies in the appearance of the same character in different images, a character locking mode is enabled, forcing the use of a fixed character reference image and seed value for generation to ensure that the appearance of the same character remains consistent across images. If the image brightness and color distribution are not up to standard, lighting constraints are added or the lighting parameters of the diffusion model are adjusted to ensure that the image brightness and color distribution meet the threshold requirements in the evaluation scheme.

[0064] If the problematic agent is an audio generation agent, the audio optimization unit is invoked for tuning. The audio optimization unit triggers corresponding optimization operations according to the pre-defined diagnostic and optimization mapping rules for the audio generation agent, based on the compliance status of each indicator in the music dimension of the evaluation scheme. If the audio-visual emotion tag matching rate is substandard, indicating that the music emotion is inconsistent with the expected scene emotion, an emotion alignment check step is added before the audio generation agent executes, forcing the generated music emotion tags to match the scene emotion tags. If the BPM and editing point alignment rate is substandard, indicating that the music beat and the visual editing are not aligned on the timeline, the editing point schedule output by the editing agent is received, forcing the music beat to bind to the editing points, synchronizing the music beat points with the visual editing points. If the volume standard deviation does not meet the requirements, indicating abnormal volume fluctuations throughout the film, an automatic volume smoothing post-processing step is added to automatically adjust the volume envelope of the generated audio, ensuring stable volume throughout the film.

[0065] If the problematic agent is the editing agent, the pacing optimization unit is invoked for tuning. Based on the compliance status of each pacing dimension indicator in the evaluation scheme, the pacing optimization unit triggers corresponding optimization operations according to the preset editing agent diagnostic and optimization mapping rules. If the number of shots in the first N seconds is insufficient, indicating inadequate opening shot density, the golden opening rule is activated, forcing a limit on the number and duration of shots in the first N seconds to ensure sufficient shot transition density in the opening phase to capture the audience's attention. If the average shot duration is insufficient, indicating overall sluggishness, long shots are automatically broken down and dynamic effects are added to improve visual pacing. If the transition type distribution is insufficient, indicating monotonous or abrupt transitions, the priority of intelligent transition strategies is increased to replace abrupt transitions and enrich the transition type distribution. If the pacing indicators still fail to meet the standards after the above operations, upstream backtracking is performed: shot-related issues are backtracked to the script generation agent to adjust the storyboard pacing and increase shot variations within the storyboard; transition issues are backtracked to the visual generation agent to add dynamic elements to the visuals, providing richer visual content support for the transitions.

[0066] Furthermore, the special dimension includes the special dimension of emotion. When the loss value of the emotion dimension is greater than the preset loss threshold, all special agents that affect the generation of the emotion curve of the film are identified and treated as problem agents. The problem type is determined to be a cross-agent problem. At this time, the problem agents include script generation agents, visual generation agents, and audio generation agents.

[0067] Reference Figure 6 The flowchart illustrating the agent tuning method for cross-agent problems demonstrates that "tuning the corresponding agent based on the determined problem type" includes the following steps: S55: When the question type is a cross-agent question, obtain the sentiment curve of the film; S56: Detect the emotional plateau interval on the emotional curve and record the start and end times and emotional values ​​of the emotional plateau interval. S57: Based on the start and end times of the emotion plain interval and the difference in emotion values ​​before and after the emotion plain interval, match the linkage instructions for each problem agent from the instruction template library. S58: Send each linkage instruction to the corresponding problem agent to perform optimization.

[0068] In S55, to address the cross-agent problem, an emotion curve optimization unit is activated. This unit slices the currently generated video into segments of fixed duration (e.g., every 2 seconds or every 5 seconds). It extracts the emotion tag sequence output by the script generation agent, the image color and brightness statistics output by the visual generation agent, and the music emotion tags and volume envelope data output by the audio generation agent within each segment. The above multimodal data is input into a pre-trained emotion analysis model for fusion and inference. The model comprehensively judges the overall emotion value conveyed by the video within each segment. The emotion value can be converted into emotion tags, such as -1 representing extremely negative, 0 representing neutral, and 1 representing extremely positive. The emotion values ​​of all segments are arranged in chronological order to form a full-film temporal emotion curve that reflects the trend of the entire film's emotion over time.

[0069] In S56, the emotion values ​​of each slice are scanned point by point along the time axis from left to right on the emotion curve. It is then checked whether there are intervals where the fluctuation range of the emotion value is consistently lower than a preset fluctuation threshold for multiple consecutive slices (e.g., the change in emotion value is less than 0.1 for more than three consecutive slices). These intervals are marked as emotion plateau intervals. For each detected emotion plateau interval, its start time (the moment the emotion enters the plateau state) and end time (the moment the emotion leaves the plateau state) are recorded, along with the emotion value sequence for each slice within that interval.

[0070] In S57, the sentiment value of the first slice after the end of the sentiment plateau interval is obtained as the post-plateau sentiment value, and the sentiment value of the last slice before the start of the sentiment plateau interval is obtained as the pre-plateau sentiment value. The difference between the two is calculated as the sentiment drop. Based on the start time of the sentiment plateau interval (which determines the time position of the instruction) and the sentiment drop (which determines the adjustment intensity of the instruction), the historical case template that most closely matches the characteristics of the current sentiment plateau interval is retrieved from the pre-built instruction template library.

[0071] The instruction template library stores successful linkage schemes for various emotional plateau scenarios. Each template records the starting time and location range of the plateau, the range of emotional drop, and the corresponding combination of linkage instruction parameters at the script, audio, and visual levels. For the script generation agent, the matched linkage instruction is to insert a plot twist before the starting time point. The intensity of the twist is determined by the specific value of the emotional drop; the larger the drop, the more intense the inserted plot twist. For the audio generation agent, the matched linkage instruction is to initiate emotional build-up before the starting time point. The build-up duration and BPM change are determined by the target emotion corresponding to the emotional value within the plateau range. For the visual generation agent, the matched linkage instruction is to initiate a color tone shift after the starting time point. The target color tone is given by the preset color scheme corresponding to the emotional value after the end of the plateau range. The specific parameters required for the above linkage instructions are obtained from the evaluation index thresholds of the emotional dimension in the evaluation scheme, the output data of each specialized agent, and the corresponding parameters of historical successful cases in the process pattern library.

[0072] In S58, each linkage instruction is sent to the corresponding target agent. For example, a plot twist insertion instruction is sent to the script generation agent. The instruction includes the specific time point of the twist insertion (a few seconds before the start time of the plateau) and the twist intensity parameter. After receiving the instruction, the script generation agent regenerates the script content at the specified position and inserts the plot twist point. An emotional build-up instruction is sent to the audio generation agent. The instruction includes the start time point of the build-up, the duration of the build-up, and the target BPM change. After receiving the instruction, the audio generation agent regenerates or adjusts the background music for the corresponding time period so that the music emotion transitions to the target emotion within the plateau interval. A color tone migration instruction is sent to the visual generation agent. The instruction includes the start time point of the color tone migration, the target color tone scheme, and the migration transition duration. After receiving the instruction, the visual generation agent regenerates the visual material for the corresponding time period so that the visual color tone gradually migrates to the target color tone after the specified time point.

[0073] After each agent completes its own re-execution, the editing agent performs a coordinated re-execution based on the new script timeline, visual materials, and audio materials, re-editing and splicing all the modified materials into a complete film. After the coordinated re-execution is completed, the film's emotional curve is regenerated, and the disappearance of the emotional plateau and the appearance of the expected peak are re-checked. If the emotional plateau disappears and the peak meets the threshold requirements of the evaluation scheme, the three-stage optimization scheme is solidified. If there are still plateau intervals, the process returns to S55 to re-analyze the current emotional curve and iterate the above process until the emotional plateau disappears and the emotional peak meets the threshold requirements of the evaluation scheme. Once the convergence condition is met, the optimized configuration is solidified.

[0074] When multiple dimensions of evaluation fail to meet the standards simultaneously, it may present both cross-agent problems and single-agent-specific problems. For example, an abnormal emotion curve and a failure to meet the script dimension may both point to the script-generating agent. In this case, simply identifying the problematic agent is insufficient to distinguish the optimization path. By identifying the problem type, it can be determined that the agent must simultaneously receive coordinated instructions from both the script optimization unit and the emotion curve optimization unit, preventing missed optimizations or optimization conflicts.

[0075] In S6, the restructured orchestration and optimized specialized agent configurations are instantiated. This involves deploying the new orchestration structure and optimized parameters for each problem agent into the generation environment, automatically triggering a new round of video generation. During the new video generation process, a new directed graph records the generation trajectory of this round. Based on the generated evaluation scheme (if the user does not provide a new round of human evaluation, the previous evaluation scheme is used; if the user provides a new evaluation, a new evaluation scheme is generated), the new video is automatically evaluated across all dimensions, outputting scores for each specialized dimension and a comprehensive evaluation score.

[0076] The evaluation results are compared with the baseline before optimization using an A / B comparison to determine whether the convergence conditions are met. Convergence conditions include achieving the target score (loss values ​​for each specific dimension and the overall evaluation score all meet the target), two consecutive rounds of positive improvement (the overall evaluation score is higher than the previous round for two consecutive rounds), and reaching the preset maximum number of iterations. If any convergence condition is met, the current round of optimization is considered successful, the optimized configuration is fixed, and the iteration ends. If none of the conditions are met, the specific agent orchestration structure detection and agent diagnosis tuning are re-executed until convergence or the preset maximum number of iterations is reached.

[0077] In S7, the final orchestration structure and optimized configuration of each specialized agent after iterative verification are synchronously updated to the current directed graph, forming a complete directed graph after this optimization. This directed graph contains both structural adjustment information and node attribute update information. This directed graph, movie type tags, loss values ​​of each specialized dimension, and comprehensive evaluation scores are organized into new historical high-quality cases and stored in the process pattern library. At the same time, all information such as problem phenomena, trajectory feature fingerprints, optimization operation records, and evaluation comparisons before and after optimization during this round of optimization are vectorized and stored in the process pattern library to provide data support for future retrieval enhancement decisions of the optimized agents.

[0078] Reference Figure 7 This disclosure provides a video generation system based on agent-based iterative optimization, comprising: The recording module 101 is used to obtain the video generation prompt words input by the user, select multiple specialized intelligent agents to generate videos based on the video generation prompt words, and use a directed graph to record the trajectory of the video generation process of multiple specialized intelligent agents. The conversion module 102 is used to obtain users' human evaluations of the film and convert the human evaluations into a structured evaluation scheme. The detection module 103 is used to detect the arrangement structure between special intelligent agents based on the current directed graph and the historical directed graph in the process pattern library when the evaluation scheme determines that the film does not meet the standards. The reconstruction module 104 is used to reconstruct the orchestration structure between special intelligent agents if the detection fails. The optimization module 105 is used to perform specific intelligent agent diagnosis based on the evaluation plan after the detection is passed, determine the problem type and the problem intelligent agent, and optimize the problem intelligent agent based on the problem type; The generation module 106 is used to regenerate the film based on the optimized special agent and the reconstructed arrangement structure until the convergence condition is met. The correction module 107 is used to correct the current directed graph based on the optimized special agent and the reconstructed orchestration structure, and store the corrected directed graph in the process pattern library.

[0079] The various variations and specific examples of the video generation method based on intelligent agent iterative optimization provided above are also applicable to the video generation system based on intelligent agent iterative optimization provided in this disclosure. Through the foregoing detailed description of the video generation method based on intelligent agent iterative optimization, those skilled in the art can clearly understand the implementation method of the video generation system based on intelligent agent iterative optimization. For the sake of brevity, it will not be described in detail here.

[0080] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0081] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the agent-based iterative optimization video generation method described in the foregoing embodiments of this disclosure.

[0082] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0083] like Figure 8 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 8 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0084] like Figure 8 As shown, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0085] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although Figure 8 A computer apparatus with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0086] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processor, all or part of the steps of the agent-based iterative optimization video generation method of embodiments of this disclosure are performed.

[0087] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0088] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the aforementioned agent-based iterative optimization-based video generation method according to embodiments of the present disclosure are performed.

[0089] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0090] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0091] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0092] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0093] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0094] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0095] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0096] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0097] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A video generation method based on agent-based iterative optimization, characterized in that, include: Obtain the video generation prompt words input by the user, select multiple specialized intelligent agents to generate videos based on the video generation prompt words, and use a directed graph to record the trajectory of the multiple specialized intelligent agents in the video generation process; Obtain user reviews of the film and transform those reviews into a structured evaluation scheme; When it is determined that the film does not meet the standard based on the evaluation scheme, the arrangement structure between special intelligent agents is detected based on the current directed graph and the historical directed graph in the process pattern library; The detection of the orchestration structure between specialized intelligent agents based on the current directed graph and historical directed graphs in the process pattern library includes: Obtain the current movie type, and filter the historical directed graph of the same movie type from the process pattern library; Obtain the structural similarity between the current directed graph and the filtered historical directed graphs; When the structural similarity is greater than a preset similarity threshold, the detection is deemed to have passed; when the structural similarity is not greater than the preset similarity threshold, the detection is deemed to have failed. If the detection fails, the orchestration structure between the specialized intelligent agents is reconstructed; The reconfiguration of the orchestration structure between the specialized intelligent agents includes: Select the historical directed graph with the highest similarity to the current directed graph structure from the process pattern library as the reference directed graph; The current directed graph is compared with the reference directed graph to identify differences. The differences include at least one of the following: missing or redundant agent nodes, differences in node execution order, inconsistencies in serial-parallel relationships, and mismatched branch conditions. Based on the differences, a target reconstruction operation is determined from multiple preset reconstruction operations. The reconstruction operation includes at least one of adding or deleting agents, adjusting the execution order, switching between serial and parallel modes, and modifying branch conditions. The target reconstruction operation is performed on the orchestration structure between the specialized intelligent agents, and the legality of the reconstructed directed graph is verified. Once the legality check passes, the reconstructed orchestration structure is re-checked. Once the check passes, the reconstructed orchestration structure is output. After the test is passed, a special intelligent agent diagnosis is performed based on the evaluation scheme to determine the problem type and the problem intelligent agent, and the problem intelligent agent is optimized based on the problem type; Based on the optimized specialized intelligent agent and the reconstructed arrangement structure, the film is regenerated until the convergence condition is met. Based on the optimized specialized intelligent agents and the reconstructed orchestration structure, the current directed graph is corrected, and the corrected directed graph is stored in the process pattern library.

2. The video generation method based on intelligent agent iterative optimization according to claim 1, characterized in that, The process of transforming the manual evaluation into a structured evaluation scheme includes: Construct an evaluation dimension table, which includes multiple special dimensions and evaluation indicators, evaluation indicator weights, evaluation indicator thresholds, and evaluation indicator collection methods for each special dimension; The manual evaluation is broken down into multiple sentences. From the evaluation dimension table, the specific dimension to which each sentence belongs is identified, and the problem description is extracted. According to the evaluation index collection method of each specific dimension, the problem description of each sentence is analyzed to obtain the corresponding evaluation index values; The evaluation index values, weights, thresholds, and data collection methods for each sentence are integrated into a structured evaluation scheme.

3. The video generation method based on intelligent agent iterative optimization according to claim 1, characterized in that, The step of performing specialized intelligent agent diagnosis based on the evaluation scheme to determine the problem type and the problem intelligent agent includes: The difference between the value of each evaluation indicator in the evaluation scheme and the corresponding evaluation indicator threshold is obtained, and the difference is weighted by the corresponding evaluation indicator weight to obtain the loss value of each special dimension. Obtain the actual output and baseline output of each specialized intelligent agent, wherein the baseline output is the historical normal output data of the same type of intelligent agent as the specialized intelligent agent extracted from the reference directed graph; Iterate through each specialized agent, replace the actual output of the currently traversed specialized agent with the baseline output, keep the links of other specialized agents unchanged, and re-obtain the loss value of the specialized dimension to which the specialized agent belongs. Use the change in the loss value before and after the replacement as the loss contribution of the specialized agent. When the loss contribution of the specialized intelligent agent is greater than the preset contribution threshold, the problem type is determined based on the type of the specialized intelligent agent, and the specialized intelligent agent is identified as the problem intelligent agent.

4. The video generation method based on agent-based iterative optimization according to claim 3, characterized in that, The specific dimension includes an emotion dimension; the step of performing specific intelligent agent diagnosis based on the evaluation scheme to determine the problem type and the problem intelligent agent also includes: When the loss value of the emotion dimension is greater than the preset loss threshold, all special agents that affect the generation of the emotion curve of the film are identified as problem agents, and the problem type is determined to be a cross-agent problem.

5. The video generation method based on agent-based iterative optimization according to claim 4, characterized in that, The optimization of the problem agent based on the problem type includes: When the problem type is a cross-agent problem, obtain the sentiment curve of the film; Detect the emotional plateau interval on the emotional curve, and record the start and end times and emotional values ​​of the emotional plateau interval; Based on the start and end times of the emotion plain interval and the difference in emotion values ​​before and after the emotion plain interval, the linkage instructions for each problem agent are matched from the instruction template library. Each linkage command is sent to the corresponding problem agent for optimization.

6. A computer device, characterized in that, The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the video generation method based on agent-based iterative optimization as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the video generation method based on agent-based iterative optimization as described in any one of claims 1-5.

8. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the steps of the video generation method based on agent-based iterative optimization as described in any one of claims 1-5.

Citation Information

Patent Citations

  • Agent cooperation-based AIGC long video stable generation method and system

    CN119893206A

  • Personnel examination answer sheet identification method based on multi-agent collaborative perception

    CN121170821A