A method and system for generating a long play based on a station-level intelligent agent arrangement

Abstract

The present application relates to the field of artificial intelligence and digital content generation technology, in particular to a kind of based on station level intelligent agent arrangement's long drama generation method and system.The method comprises: automatically completing production link according to input type by distribution agent, scheduling workshop intelligent agent to execute task, and workshop intelligent agent is divided into multiple stations inside;Full-link mapping table from video frame to source station is constructed synchronously in production process;Support independent modification to the output data of any station and only trigger the incremental recalculation of downstream station;When generating video script or video frame, asset attribute is read from asset knowledge graph and forced injection is generated to ensure visual consistency;In response to user's specified operation on video defect, based on full-link mapping table, locate to source station in reverse and repair and automatically suture at fixed point.The present application realizes the fine control, reversible repair and consistency guarantee of AI video generation process, significantly reduces the power consumption, and improves the production efficiency and the quality of the film.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and digital content generation technology, specifically to a method and system for generating comics based on workstation-level intelligent agent arrangement. Background Technology

[0002] With the rapid development of AIGC (Artificial Intelligence Generated Content) technology, video generation technology based on large language models and diffusion models has been initially applied to the production of digital content such as comics and short dramas. Currently, the mainstream AI video generation solutions in the industry mainly include the following categories: (1) Single-model end-to-end generation scheme: The user inputs a text description or reference image, and the large model directly outputs a complete video clip. Typical examples include Runway Gen series and Pika Labs. This scheme generates videos quickly, but it has a "black box" characteristic. Users cannot intervene in or control the details of the generation process. When the generated result is flawed, the input parameters can only be adjusted and the entire process can be regenerated, which is inefficient.

[0003] (2) Agent pipeline solution: Some platforms (such as iQiyi NaDou Pro and Beijing Baidu Netcom Technology Co., Ltd.'s CN121691839A "Video Generation Method, Device and System Based on Multi-Agent Collaboration") propose to decompose script breakdown, storyboard generation, video synthesis and other links into multiple independent agents, which are executed serially in a fixed order. Compared with the end-to-end solution, the agent solution improves the controllability of the process, but its granularity is still at the "agent" level. Each agent runs as a whole, and the internal process is still invisible and cannot be interfered with. Although the Suzhou Daily's CN120547420B "An AI Intelligent Short Video Generation Method and System Based on Multi-Agent Collaboration" introduces a unified picture style algorithm, this algorithm is a passive adjustment in the post-processing stage and cannot prevent visual inconsistency problems from the source. The CN121056706A "A Video Generation Method and Device Based on Multi-Agent Collaboration" published by Wuhan Sinan Yiyi Intelligent Technology Co., Ltd. implements the function of selecting intelligent agents according to input type, but the process sequence is fixed and lacks adaptive completion and reverse reasoning capabilities.

[0004] (3) Video defect repair solutions: Companies such as Danghong Technology have proposed video defect removal methods based on deep learning, which automatically detect and repair video defects through motion information estimation. This method belongs to the category of "post-repair", and its repair capability is limited by the model performance. It cannot solve systemic defects caused by upstream script or storyboard design problems.

[0005] A search revealed that the closest existing technology to this solution is iQiyi's publicly disclosed AI-powered anime production agent cluster solution. This solution proposes decomposing the anime production process into multiple agent-driven stages, achieving full automation from script to finished product. However, this solution remains at the agent level in terms of control granularity, failing to support step-by-step modification of intermediate outputs within a single agent; it lacks a systematic tracing and repair mechanism for defect handling; it employs image space constraints via subject library / standard character graph replacement for consistency assurance, rather than structured semantic constraints; and it only supports unidirectional generation in terms of process direction, unable to automatically adapt or reverse-adjust the production process based on input type.

[0006] In summary, existing technologies generally suffer from the following technical problems: 1. Coarse-grained and lacking fine-grained controllability. The granularity of the intelligent agent is indivisible, the intermediate processes are invisible, and users cannot precisely intervene in specific stages. When a user wants to adjust a minor step in the generation process, the entire intelligent agent or even the entire process must be re-executed, resulting in a serious waste of computing power.

[0007] 2. The generated results are irreversible, and the cost of repairing defects is high. Video generation is a unidirectional pipeline; once a defect appears in the finished product, it is impossible to trace back to the specific step and parameters that caused the defect, and the entire process must be regenerated. Existing consistency checks can only detect problems but do not have the ability to locate and repair them.

[0008] 3. Consistency of assets is difficult to guarantee. The visual characteristics of the same character are inconsistent in different shots (giveaways). Existing solutions either use image feature retrieval (which is susceptible to the randomness of the generative model) or post-production style unification algorithms (which can only be passively adjusted), lacking an effective mechanism for explicit constraints from the source.

[0009] 4. Rigid processes and poor adaptability. The system cannot automatically adapt to the production process based on user input types, nor can it support reverse production links (such as deducing storyboards or scripts from existing videos), which limits the system's applicability and user experience. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention provides a method and system for generating animated series based on workstation-level intelligent agent orchestration. By organically combining technical features such as workstation-level independent control and incremental recalculation mechanism, full-link reverse tracing and fixed-point repair, semantic-level asset consistency constraints based on knowledge graph, and adaptive bidirectional process orchestration, it solves the core pain points of existing AI video generation methods, such as poor controllability, poor reversibility, poor consistency, and rigid processes. Significant technological progress has been achieved in reducing computing power consumption, improving production efficiency, ensuring the quality of finished products, and flexibly adapting to needs.

[0011] This invention is achieved through the following technical solution: A method for generating animated series based on workstation-level intelligent agent orchestration is provided, including the following steps: S1. Receive user input through a distribution agent, automatically determine and complete the required production link based on the type of user input, and schedule the production line composed of multiple workshop agents to execute tasks; each workshop agent is internally divided into multiple workstations, each workstation has an independent data structure, status management, and version snapshot. The data structure includes at least workstation number, workstation name, status, input data, output data, and version history. The status includes at least: draft, processing, confirmed, and locked. S2. During the forward execution of tasks on the production line, a full-link mapping table is simultaneously constructed, mapping from the final generated video frames to storyboard shots, assets, scripts, and even the source workstations, and the input / output data and version information of each workstation are recorded. S3. In response to a user's request to modify the output data of any workstation, verify that the format of the modified data conforms to the predefined data contract of the workstation, save the current version snapshot and increment the version number, write the modified data to the workstation, and then only trigger the downstream workstations that depend on the modified workstation to perform incremental recalculation, while keeping the data of upstream and unrelated workstations unchanged. S4. When generating video storyboards or video frames, read the attribute descriptions of relevant assets from the pre-built asset knowledge graph and use them as constraints for the generation model to forcefully inject generation prompts, ensuring the consistency of asset visual features across multiple shots. The asset knowledge graph includes character nodes, scene nodes, and prop nodes, as well as edges describing the relationships between nodes. Each node stores quantifiable visual attribute labels. S5. In response to the user's specified operation on the defects in the video, the source station and its parameters are located in reverse based on the full-link mapping table, and an editing interface for modifying the parameters of the source station is provided. S6. After receiving the modification of the source station parameters, only the incremental recalculation of the downstream stations starting from the source station is triggered. The regenerated content is automatically stitched with the unaffected parts of the original video, and the repaired video is output.

[0012] In this scheme, the relationship between adjacent steps S1-S6 is not a simple linear sequence, but rather includes various logical relationships such as temporal sequence, parallel synchronization, and conditional triggering. Specifically: S1 is the starting point of the process: the agent receives input, determines the type, completes the link, and schedules the pipeline to start execution. S2 is performed synchronously during the "forward execution of tasks on the production pipeline" initiated by S1. S2 immediately begins to construct the mapping table in parallel after the pipeline starts execution.

[0013] S2 continuously records the mapping table and is a background process; S3 responds to user modification requests for any workstation, which can be triggered at any time during pipeline execution. S3 does not depend on the completion of S2, and the two can occur concurrently. Modifications to S3 trigger incremental recalculation, while S2 continues to update the mapping table during the recalculation process. S2 and S3 are asynchronous and parallel.

[0014] S3 is a user-initiated modification that can occur at any stage of the pipeline execution.

[0015] S4 is the knowledge graph constraint injection that is automatically executed during the generation of storyboard agents and video agents in the normal execution of the pipeline. S4 is an inherent step in the production process, while S3 is an optional user intervention. If the user modifies the upstream workstation before S4 is executed, S4 will execute based on the modified data; if the user modifies after S4 is executed, the output already completed by S4 may be overwritten by subsequent incremental recalculation.

[0016] S4 generates video storyboards and video frames, which are the steps in producing video content.

[0017] S5 responds to user-specified actions regarding defects in the video, provided the video has already been generated (i.e., S4 has completed). Only after S4 has finished executing and outputting the video can the user see the defects and perform a double-click operation.

[0018] The S5 completes the reverse positioning and displays the editing interface. The user modifies the parameters and submits.

[0019] S6 receives the modified parameters, triggering incremental recalculation and stitching output. After S5 completes (and the user confirms the modification), S6 is executed immediately. The two are strictly sequential in timing.

[0020] This diverse step-by-step design enables the system to automatically produce according to the standard process (S1-S2-S4), and to respond to user-initiated modifications (S3) and defect repairs (S5-S6) at any time, demonstrating the flexibility and robustness of the invention.

[0021] Furthermore, in S1, the required production chain is automatically determined and completed based on the type of user input. Specifically, if the user input is a single sentence of text or a single image, the material agent is automatically invoked to expand or understand it, and the complete workshop agent process is completed. If the user input is a video clip, the process starts from the storyboard agent or resource agent and works backward to complete the script or material information. If the user input is a long text or script document, all workshop agents are scheduled in a standard order.

[0022] Furthermore, in S3 or S6, incremental recalculation is triggered, which specifically includes: identifying the downstream dependent station chain from the modified source station to the final video output; performing recalculation on the stations in the downstream dependent station chain in sequence, wherein the recalculation of each station only uses the latest output data of its upstream station.

[0023] Furthermore, the method also includes a consistency check step: extracting the visual attribute features of the actual characters from the generated video clips; obtaining the standard features of the corresponding characters from the asset knowledge graph; calculating the similarity between the actual features and the standard features; if the similarity is lower than a preset threshold, a consistency violation is recorded and the user is notified to make corrections.

[0024] Furthermore, in S4, the forced injection of generated prompts also includes: when the storyboard agent generates prompts for the first and last frames, automatically reading the character's appearance tags and clothing style from the asset knowledge graph and using them as part of the positive prompts; at the same time, generating negative prompts to prevent the generation model from changing the character's core visual attributes; and after the video agent generates video frames, verifying the consistency between the character's visual features in the generated results and the attributes described in the asset knowledge graph.

[0025] Furthermore, the range locking method for incremental recalculation in S6 includes: Vertical locking: Only the downstream dependency chain from the modified source station to the final video output is re-executed; Lateral Lock: Only regenerates the shots or clips affected by the modification; adjacent unaffected shots remain unchanged.

[0026] Furthermore, it also includes asset locking steps: When any asset in the asset knowledge graph is confirmed to be locked, its core attributes are marked as unmodifiable. Downstream generation steps must not change the locked attributes of the asset when generating prompts or video frames, and the system automatically verifies and prevents any operation that attempts to modify the locked attributes during the generation process.

[0027] This invention also provides a comic book generation system using a comic book generation method based on workstation-level intelligent agent arrangement, comprising: Distribute intelligent agents to enable S1 to automatically determine and complete the required production links based on the type of user input, and schedule the production line consisting of multiple workshop intelligent agents to execute tasks. The workstation-level control module is used to divide each workshop agent into multiple workstations in S1, and to enable independent modification of the output data of any workstation in S3, and to trigger incremental recalculation only for downstream workstations that depend on the modified workstation, while keeping the data of upstream and unrelated workstations unchanged. The forward generation and reverse tracing modules are used to simultaneously build a full-link mapping table during the forward execution of tasks on the production line in S2; to respond to user-specified operations on defects in the video in S5, and to reverse locate the source station and its parameters based on the full-link mapping table; and to provide an editing interface for modifying the parameters of the source station in S6, and to trigger incremental recalculation starting from the source station after receiving the modification, automatically stitching the regenerated content with the unaffected parts of the original video, and outputting the repaired video. The asset knowledge graph module stores character nodes, scene nodes, and prop nodes, as well as edges describing the relationships between nodes. Each node stores quantifiable visual attribute labels. It is used in S4 to read the attribute descriptions of relevant assets from the asset knowledge graph when generating video storyboards or video frames, and to forcefully inject them into the generation prompts as constraints of the generation model to ensure the consistency of the visual features of assets across multiple shots.

[0028] Furthermore, the distribution agents are configured as follows: if the user input is a single sentence of text or a single image, the material agent is automatically invoked to expand or understand it, and complete the entire workshop agent process; if the user input is a video clip, the process starts from the storyboard agent or resource agent and works backward to complete the script or material information; if the user input is a long text or script document, all workshop agents are scheduled in a standard order.

[0029] The beneficial effects of this invention are: This invention achieves the following through innovative mechanisms: workstation-level independent control and incremental recalculation, end-to-end reverse tracing and pinpoint repair, knowledge graph-based semantic-level asset consistency constraints, and adaptive bidirectional process orchestration. I. Achieve refined and controllable generation process, significantly reducing computing power consumption. This invention splits each workshop agent into multiple stations with independent data structures, state management, and version snapshots, and supports independent modification of the output data of any station. Only downstream stations that depend on the modified station are triggered to perform incremental recalculation, while upstream and unrelated stations remain unchanged.

[0030] Users can precisely intervene and modify any fine-grained step in the production pipeline, completely breaking the limitation of the uncontrollable black box at the intelligent agent level in existing technologies, and realizing a non-destructive editing experience similar to professional creative software (such as Photoshop and Premiere).

[0031] The incremental recalculation mechanism avoids the waste of computing power caused by triggering a complete regeneration due to minor modifications. Experimental data shows that, in the generation of videos of the same quality, the large model's word consumption is reduced by more than 40%, significantly reducing computing power costs and generation latency.

[0032] II. Establishing a closed loop for defect tracing and targeted repair greatly improves production efficiency. This invention synchronously constructs a full-link mapping table during the forward production process, mapping from the final generated video frames to storyboard shots, assets, scripts, and even the source workstations. When a user double-clicks to specify a flaw in the video, the system can reverse-engineer to accurately locate the source workstation and its parameters that caused the flaw. After the user makes a modification, only a local incremental recalculation is triggered, and the repaired content is automatically stitched together with the unaffected parts of the original video.

[0033] This reduces the closed-loop time of several hours in traditional solutions, where the entire video must be regenerated after a flaw is discovered, to just minutes. It enables highly efficient iteration from flaw discovery to source identification, parameter modification, and output of the repaired video. The targeted repair mechanism eliminates the need for users to regenerate the entire video, significantly reducing waiting time and trial-and-error costs, making it particularly suitable for large-scale production scenarios such as long videos and multi-episode animated series.

[0034] III. Ensuring visual consistency across shots based on semantic-level constraints using knowledge graphs This invention constructs an asset knowledge graph containing three types of nodes: characters, scenes, and props, as well as relational edges. Each node stores quantifiable visual attribute tags (such as appearance tags, clothing style, hair color, etc.). During the storyboard prompt generation and video generation stages, the system forcibly reads attributes from the graph and injects them into the generated prompts (including positive constraints and negative prohibitions), and performs an algorithmic consistency check after the final cut.

[0035] This solves the common visual inconsistencies in AI-generated long videos, such as character "bloopers" and scene "collapses," significantly improving the professionalism and watchability of the finished product.

[0036] Unlike existing technologies that rely on passive solutions such as image feature retrieval or post-processing style unification, the semantic-level explicit constraint mechanism of this invention is more stable and reliable, and supports asset locking functionality to ensure that the confirmed core attributes are not modified during the downstream generation process.

[0037] IV. Adaptive process completion and bidirectional production, flexibly adapting to diverse needs. This invention supports automatically identifying and completing the required production chain based on the user input type (single sentence text, single image, video clip, long text / script), and also supports reverse reasoning from existing video clips to generate upstream content such as storyboards and scripts.

[0038] Users no longer need to manually configure complex production processes; the system automatically adapts to the optimal production path, lowering the barrier to entry and enabling even non-professional users to easily generate high-quality comics. The reverse production mode provides a new approach for secondary content creation, adaptation, and material reuse, expanding the system's application scenarios. It supports user-defined selection of specific intelligent agents for segmented generation, with the system automatically filling in any missing intermediate steps, balancing flexibility and process integrity.

[0039] V. Synergistic effects of multiple features to build a complete creative iteration ecosystem The four core innovations mentioned above are not isolated but interdependent and synergistic. Workstation-level control provides an operable workstation granularity for reverse traceability, the full-link mapping table provides a data foundation for reverse localization, knowledge graph constraints reduce the probability of defects from the source, and incremental recalculation makes modification costs extremely low.

[0040] The system forms a complete closed loop of "creation-preview-problem discovery-source identification-parameter modification-partial recalculation-output of optimized finished product", which greatly improves the usability and user satisfaction of the AI ​​video generation system.

[0041] This technical solution can be widely used in film and television production, short video creation, advertising and marketing, education and training, and other fields, and has significant commercial value and broad application prospects. Attached Figure Description

[0042] Figure 1 This is a diagram of the overall system architecture of the present invention.

[0043] Figure 2 This is a flowchart of the workstation-level intelligent agent orchestration process in this invention.

[0044] Figure 3 This is a flowchart of the defect tracing and fixed-point regeneration process in this invention.

[0045] Figure 4 This is a schematic diagram of the asset map structure in this invention.

[0046] Figure 5 This is a flowchart of the intelligent process completion method in this invention. Detailed Implementation

[0047] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to describe the solution.

[0048] Example 1:

[0049] This embodiment provides a specific implementation of a comic book generation system based on workstation-level intelligent agent arrangement.

[0050] (I) System Initialization When the system starts up, such as Figure 1 and Figure 2 As shown, the workstation definitions of all workshop agents are loaded first. Taking the material agent as an example, this agent is divided into multiple workstations. Each workstation contains a unique workstation number, workstation name, corresponding processor function, and optional large language model configuration. Specifically, the material agent includes the following workstations: workstation 1-1-1 is responsible for material uploading, workstation 1-1-2 is responsible for format verification, workstation 1-1-3 is responsible for content extraction (this workstation calls the Hunyuan large model for processing), workstation 1-2-1 is responsible for diversity recognition, workstation 1-2-2 is responsible for character extraction, workstation 1-2-3 is responsible for prop recognition, workstation 1-2-4 is responsible for scene recognition, and workstation 1-3-4 is responsible for quality inspection report generation. The workstations communicate with each other through predefined data contracts to ensure the consistency of upstream and downstream data formats. The data structure of each workstation also includes a "locker identifier" field, used to record the user or process identifier currently locking the workstation to prevent data conflicts caused by multiple users operating simultaneously.

[0051] (ii) Task submission and workstation execution When a user submits a production task, the distribution agent automatically generates an execution plan based on the type and content of the user's input. For example, when a user uploads a novel PDF file and specifies the generation of a 10-episode comic series, each episode being 150 seconds long, the distribution agent will schedule the material agent, script agent, resource agent, storyboard agent, and video agent to execute in sequence.

[0052] During the process, a status snapshot is automatically saved after each workstation completes its processing.

[0053] The specific process is as follows: First, the workstation status is set to "Processing". Then, the input data is written to the workstation and a snapshot is saved. Next, the processor function of that workstation is called to generate output data. Finally, the output data is written to the workstation, the status is updated to "Confirmed", the version number is incremented, and a snapshot is saved again. The entire process ensures that the intermediate results of each workstation are traceable and rollbackable.

[0054] (III) Example of workstation data modification When users find errors in the diversity identification results, they can directly modify the output data of workstation 1-2-1 (diversity identification) through the system interface. Users only need to specify the target workstation ID and the new output data (e.g., a redefined list of chapters), and the system will automatically perform the following operations: First, verify whether the format of the modified data conforms to the predefined data contract of the workstation. Then, save the current workstation data to the version history and increment the version number by 1. Next, write the modified data to the current workstation. Finally, identify the downstream dependent workstation chain from the workstation to the final output and perform recalculation on the downstream workstations in sequence. Each downstream workstation only uses the latest output data of its upstream workstation.

[0055] In this example, the system only re-executes workstations 1-2-2 (character extraction), 1-2-3 (prop recognition), 1-2-4 (scene recognition), and all subsequent workstations, while the upstream workstations of the material agent (1-1-1 to 1-1-3) remain unchanged, thus achieving precise local modification.

[0056] Performance Verification: Taking the modification of the episode recognition workstation (1-2-1) as an example, in a comic strip generation task containing 10 episodes, each 150 seconds long, if the traditional overall recalculation scheme is used, each modification requires re-executing all 5 workshop agents, taking an average of about 45 minutes and consuming about 1.2 million tokens. However, using the workstation-level incremental recalculation mechanism of this invention, only workstations 1-2-2, 1-2-3, 1-2-4, and subsequent downstream workstations need to be re-executed, taking an average of about 8 minutes and consuming about 700,000 tokens. In this example scenario, the number of large model calls is reduced by about 75%, and token consumption is reduced by more than 40%. This advantage becomes more significant as the video length increases.

[0057] Example 2:

[0058] This embodiment provides a specific implementation of a defect tracing and targeted repair method based on end-to-end mapping.

[0059] (a) Establishing a mapping table During the production process, the video agent is responsible for recording the mapping relationship between each video frame and its source information. Specifically, the system maintains a mapping table, where each entry contains the video ID, frame number, corresponding storyboard ID, source workstation identifier, and source data content. Once the video is generated, this mapping table records the complete link from each frame to its source.

[0060] (II) Front-end double-click interaction and back-end tracing like Figure 3 As shown, a defect capture event is embedded in the front-end player. When a user double-clicks on a defective area in the video frame, the front-end obtains the normalized coordinates of the mouse click location and the frame number corresponding to the current playback time, and then sends the video ID, frame number, and coordinate information to the back-end service. Upon receiving the request, the back-end service queries the mapping table for the source information corresponding to that frame. If the exact frame is not recorded, it searches for the mapping of the nearest frame as an approximate result. After the query is complete, the back-end returns the source workstation identifier, source data content, and repair suggestions. The front-end then pops up the source workstation's editing panel for the user to view and modify relevant parameters.

[0061] (III) Fixed-point regeneration After the user confirms the changes, the system executes the incremental regeneration process.

[0062] First, the system calculates the range of affected segments (i.e., the video segments involved in the downstream link from the source station to the final output) based on the source station identifier and task ID.

[0063] Then, only the affected segments are regenerated using the modified parameters, while the unaffected segments remain unchanged.

[0064] Finally, the system performs frame-level alignment and splicing between the regenerated segments and the unaffected parts of the original video, outputting the repaired complete video.

[0065] To achieve precise local repair, this invention employs a range locking method, including longitudinal locking and lateral locking: 1. Vertical Locking: The system only re-executes the downstream dependent workstations in the chain from the modified source workstation to the final video output, while upstream workstations and parallel workstations unrelated to the modification remain unchanged. For example, if the "asset_id generation" workstation in the resource agent is modified, only its downstream workstations such as "map construction," "storyboard generation," and "video generation" are re-executed, while all workstations in the material agent and script agent remain unaffected.

[0066] 2. Lateral Locking: At the video segment level, the system only regenerates the shots or segments affected by the modification, while adjacent unaffected shots remain unchanged. The system accurately calculates the range of shots affected by the modification through a full-link mapping table, avoiding unnecessary regeneration.

[0067] Performance Verification: For a 10-minute video clip, traditional methods require complete regeneration after a flaw is detected, taking an average of about 2 hours (including regeneration and manual inspection). This invention, through reverse localization and pinpoint repair mechanisms, reduces the average closed-loop time from the user double-clicking on a flaw to outputting the repaired video to 4.5 minutes, improving efficiency by more than 25 times. In 50 tests, the flaw localization accuracy reached 93%, and the smoothness of the transition between the repaired video and the unaffected portion scored 4.6 out of 5.

[0068] Example 3:

[0069] This embodiment provides a specific implementation of an asset consistency assurance method based on knowledge graphs.

[0070] (I) Graph Data Structure like Figure 4As shown, the system uses a graph database (such as Neo4j) to store the asset knowledge graph. The graph contains three types of nodes: character nodes, scene nodes, and item nodes. Each character node stores attributes such as asset ID, name, appearance tags (e.g., "ancient costume," "long hair," "swordsman," "white clothes"), clothing style, voice ID, height, and hair color. Each scene node stores attributes such as asset ID, location name, atmosphere description, and historical background. Each item node stores attributes such as asset ID, name, and visual description. In addition, the graph also includes edges describing the relationships between nodes, such as the "appearance" relationship between a character and a scene (which can include frequency and importance levels), the "use" relationship between a character and an item, and the "interaction" relationship between characters.

[0071] (ii) Graph Constraints During Prompt Word Generation When the storyboard agent generates prompts for the first and last frames, the system automatically reads the standard descriptions of the relevant characters from the asset knowledge graph. Specifically, based on the character asset ID in the storyboard information, the system retrieves the character's appearance tags, clothing style, and other attributes from the graph, then forcibly injects these attributes into the positive prompts, while simultaneously generating negative prompts to prevent the generation model from altering the character's core visual features. For example, for the character "Zhang San," the system will generate positive prompts including "ancient costume, long hair, swordsman, white-clothes style, chivalrous customer service, maintain Zhang San's consistent appearance, do not change core features such as clothing, hairstyle, and accessories, cinematic quality," and negative prompts including "character deformation, clothing change, hairstyle change, multiple limbs, deformity, low resolution." This method ensures that the character's visual features in each shot remain consistent with the standard descriptions in the graph.

[0072] (iii) Consistency check After the final video clips are generated, the system performs a consistency check. For each character appearing in each video clip, the system first uses a computer vision model to extract the character's actual visual attribute features from the video frames, then retrieves the character's standard features from the asset knowledge graph, and finally calculates the cosine similarity between the two. If the similarity is lower than a preset threshold (e.g., 0.7), a consistency violation is recorded, including the character name, clip ID, similarity score, and modification suggestions (e.g., "Check if the generation prompt for this character contains all appearance tags"). The system summarizes all violation information and notifies the user, who can then modify the relevant workstations according to the suggestions and trigger an incremental recalculation.

[0073] To further ensure asset consistency, this invention supports an asset locking function. Once the core attributes of an asset (such as the character "Zhang San"), such as its appearance tags and clothing style, are confirmed by the user, the asset's status can be set to "locked." After locking, the asset's core attributes are marked as unmodifiable, and no downstream generation step (including the generation of prompts by the storyboard agent and the generation of video frames by the video agent) may change the locked attributes. The system automatically verifies during the generation process. If it detects that the generated model is attempting to output features that do not conform to the locked attributes, it will forcibly correct or refuse the output and issue an alarm. This mechanism eliminates the asset attribute drift problem caused by the randomness of downstream generation at the institutional level.

[0074] Performance Verification: In a test scenario comprising 20 shots and 5 characters, without knowledge graph constraints, the average score for visual feature consistency was 0.52 (cosine similarity), with a detection rate of 34% (i.e., the percentage of visual inconsistencies appearing in the 20 shots). After applying the knowledge graph constraints of this invention, the average consistency score increased to 0.89, and the detection rate decreased to 3.2%. The asset locking mechanism further ensures zero modification of core attributes during downstream generation. In 50 repeated tests, the performance remained stable with a standard deviation of less than 0.05.

[0075] Example 4:

[0076] This embodiment provides a specific implementation of an intelligent process completion method based on input type.

[0077] (a) Input type judgment like Figure 5 As shown, the system first performs a type determination on the user's input. The determination logic is as follows: If the input is plain text and has fewer than 500 characters, it is considered short text; if the input is plain text and has 500 or more characters, it is considered long text (novel / script); if the input begins with an image data URI or ends with a common image format suffix, it is considered an image; if the input begins with a video data URI or ends with a common video format suffix, it is considered a video; if the input ends with a document format suffix such as PDF, DOCX, or TXT, it is considered a document; otherwise, it is considered an unknown type.

[0078] (II) Process Completion Algorithm The system automatically completes the production process based on the agent selected by the user and the input type. The standard workflow sequence is: source agent, script agent, resource agent, storyboard agent, and video agent. If the user does not select any agent (one-click completion mode), the starting point is determined based on the input type: for short text, images, or documents, a complete forward five-workshop workflow is used. For videos, a reverse process is used (reasoning backwards from the video agent). If the user selects only some agents (segmented generation mode), the system will automatically fill in the missing intermediate links between the selected agents, forming a continuous process from the smallest number to the largest number.

[0079] (III) Execution Plan Generation The system generates a detailed execution plan based on the completed process chain. The execution plan includes the execution chain (a sequentially arranged list of agents), dependencies (previous agents and their data contracts that each agent depends on), estimated execution time, and input preparation information. For the first agent in the process chain, the system performs necessary format conversions based on the type of user input to ensure that the input data conforms to the agent's data contract requirements. After the execution plan is generated, it is submitted to the scheduling engine for sequential execution.

[0080] (iv) Implementation of the reverse process When a user uploads an existing video clip and wishes to reverse engineer it into a storyboard or script, the system executes a reverse engineering process. The specific steps are as follows: First, the system calls the reverse reasoning interface of the video agent to extract keyframe sequences from the video; second, it calls the resource agent to infer scene and character information based on the keyframes, identifying assets appearing in the video and their attributes; then, it calls the script agent to infer the script structure based on the scene and character sequences, generating episode outlines and dialogue drafts; finally, it outputs semi-structured storyboards and scripts for users to edit and modify, or as a starting point for new comic book projects. This reverse engineering process allows users to quickly create or adapt content based on existing video materials.

[0081] Performance Verification: The system supports five input types (short text, long text, image, video, and document). In one-click generation mode, users can automatically adapt to the complete workflow without any configuration. In 100 tests, the input type recognition accuracy reached 98%, the workflow completion accuracy was 96%, and the reverse reasoning success rate (deriving storyboards from video) was 87%. For segmented generation mode, when the user selects only some agents, the system's success rate in automatically filling in missing intermediate links was 94%.

[0082] The four embodiments above correspond to the four core innovations of this invention: system architecture, defect repair, consistency assurance, and process completion, comprehensively demonstrating the specific implementation of this invention. These four embodiments are interdependent and synergistic, together constituting the complete technical solution of this invention. Workstation-level fine control (Example 1) provides an operable workstation granularity for reverse tracing (Example 2). Without workstation-level division, reverse positioning can only reach the agent level, and cannot accurately locate the specific link and parameters that cause defects, thus making efficient point-to-point repair impossible.

[0083] The end-to-end mapping table (Example 2) provides the data infrastructure for reverse localization. Without synchronously building the mapping relationship from video frames to the source workstation during the forward generation process, even if the system has workstation-level control capabilities, users cannot trace back to the specific modification location from the defect in the finished product.

[0084] Knowledge graph-based asset consistency constraints (Example 3) reduce the probability of defects from the source. By forcibly injecting semantic constraints, issues such as character inconsistencies and scene breakdowns are suppressed during the generation stage, thereby reducing the frequency of reverse tracing and targeted repairs and improving overall production efficiency.

[0085] The incremental recalculation mechanism (used throughout Examples 1 and 2) makes the cost of repairs extremely low. Whether it's a general modification or a defect repair scenario, the system only performs local recalculation on the affected downstream workstations, rather than regenerating the entire system. This greatly improves users' willingness and efficiency in iterative modifications.

[0086] The above four features form a complete creative iteration loop of "creation-preview-problem discovery-source location-parameter modification-local recalculation-output of optimized finished product", which enables the present invention to have non-destructive editing and local rendering capabilities similar to professional creation software (such as Photoshop and Premiere), which is unprecedented in the field of AI video generation.

[0087] Of course, the above description is not limited to the examples above. Technical features not described in this invention can be implemented by or using existing technology, and will not be repeated here. The above embodiments are only used to illustrate the technical solutions of this invention and are not intended to limit this invention. This invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention do not depart from the spirit of this invention and should also fall within the scope of protection of the claims of this invention.

Claims

1. A method for generating animated series based on workstation-level intelligent agent arrangement, characterized in that: Includes the following steps: S1. Receive user input through a distribution agent, automatically determine and complete the required production link based on the type of user input, and schedule the production line composed of multiple workshop agents to execute tasks; each workshop agent is internally divided into multiple workstations, each workstation has an independent data structure, status management, and version snapshot. The data structure includes at least workstation number, workstation name, status, input data, output data, and version history. The status includes at least: draft, processing, confirmed, and locked. S2. During the forward execution of tasks on the production line, a full-link mapping table is simultaneously constructed, mapping from the final generated video frames to storyboard shots, assets, scripts, and even the source workstations, and the input / output data and version information of each workstation are recorded. S3. In response to a user's request to modify the output data of any workstation, verify that the format of the modified data conforms to the predefined data contract of the workstation, save the current version snapshot and increment the version number, write the modified data to the workstation, and then only trigger the downstream workstations that depend on the modified workstation to perform incremental recalculation, while keeping the data of upstream and unrelated workstations unchanged. S4. When generating video storyboards or video frames, read the attribute descriptions of relevant assets from the pre-built asset knowledge graph and use them as constraints for the generation model to forcefully inject generation prompts, ensuring the consistency of asset visual features across multiple shots. The asset knowledge graph includes character nodes, scene nodes, and prop nodes, as well as edges describing the relationships between nodes. Each node stores quantifiable visual attribute labels. S5. In response to the user's specified operation on the defects in the video, the source station and its parameters are located in reverse based on the full-link mapping table, and an editing interface for modifying the parameters of the source station is provided. S6. After receiving the modification of the source station parameters, only the incremental recalculation of the downstream stations starting from the source station is triggered. The regenerated content is automatically stitched with the unaffected parts of the original video, and the repaired video is output.

2. The method for generating comics based on workstation-level intelligent agent arrangement according to claim 1, characterized in that: In S1, the required production link is automatically determined and completed based on the type of user input, specifically including: If the user inputs a single sentence of text or a single image, the intelligent agent will be automatically invoked to expand or understand the material and complete the entire workshop intelligent agent process. If the user input is a video clip, then reverse reasoning starts from the storyboard agent or resource agent to complete the script or material information in the forward direction; If the user input is a long text or a script document, all workshop agents are scheduled in the standard order.

3. The method for generating comics based on workstation-level intelligent agent arrangement according to claim 1, characterized in that: In S3 or S6, incremental recalculation is triggered, specifically including: Identify the downstream dependent workstation chain from the modified source workstation to the final video output; Recalculation is performed sequentially on each workstation in the downstream dependent workstation chain, where the recalculation of each workstation uses only the latest output data of its upstream workstation.

4. The method for generating comics based on workstation-level intelligent agent arrangement according to claim 1, characterized in that: It also includes a consistency check step: Extract the visual attribute features of the actual characters appearing in the generated video clips; Obtain the standard characteristics of the corresponding roles from the asset knowledge graph; Calculate the similarity between actual features and standard features; If the similarity is lower than a preset threshold, a consistency violation will be recorded, and the user will be notified to make corrections.

5. The method for generating comics based on workstation-level intelligent agent arrangement according to claim 1, characterized in that: In S4, forced injection of prompt words also includes: When the storyboard agent generates prompts for the first and last frames, it automatically reads the character's appearance tags and clothing style from the asset knowledge graph and uses them as part of the positive prompts. At the same time, negative prompts are generated to prevent the generated model from altering the character's core visual attributes; After the video agent generates video frames, verify the consistency between the visual features of the characters in the generated results and the attributes described in the asset knowledge graph.

6. The method for generating comics based on workstation-level intelligent agent arrangement according to claim 1, characterized in that: In S6, the range locking methods for incremental recalculation include: Vertical locking: Only the downstream dependency chain from the modified source station to the final video output is re-executed; Lateral Lock: Only regenerates the shots or clips affected by the modification; adjacent unaffected shots remain unchanged.

7. The method for generating comics based on workstation-level intelligent agent arrangement according to claim 1, characterized in that: It also includes asset locking steps: When any asset in the asset knowledge graph is confirmed to be locked, its core attributes are marked as unmodifiable. Downstream generation steps must not change the locked attributes of the asset when generating prompts or video frames, and the system automatically verifies and prevents any operation that attempts to modify the locked attributes during the generation process.

8. A comic strip generation system using the comic strip generation method based on workstation-level intelligent agent arrangement as described in any one of claims 1 to 7, characterized in that: include: Distributing intelligent agents: used to implement in S1: automatically determine and complete the required production links based on the type of user input, and schedule the production line composed of multiple workshop intelligent agents to execute tasks; Workstation-level control module: used to implement S1: dividing each workshop agent into multiple workstations; and used to implement S3: supporting independent modification of the output data of any workstation, and only triggering incremental recalculation of downstream workstations that depend on the modified workstation, while keeping the data of upstream and unrelated workstations unchanged; The forward generation and reverse tracing module is used to implement the following in S2: During the forward execution of tasks on the production line, a full-link mapping table is simultaneously constructed, mapping from the final generated video frames to storyboard shots, assets, scripts, and even the source workstations. It is also used in S5: In response to user-specified operations on video defects, the module reversely locates the source workstation and its parameters based on the full-link mapping table. Finally, it is used in S6: Providing an editing interface for modifying source workstation parameters, and upon receiving modifications, triggering incremental recalculation starting from the source workstation, automatically stitching the regenerated content with the unaffected parts of the original video, and outputting the repaired video. The asset knowledge graph module stores character nodes, scene nodes, and prop nodes, as well as edges describing the relationships between nodes. Each node stores quantifiable visual attribute labels. In S4, when generating video storyboards or video frames, the attribute descriptions of relevant assets are read from the asset knowledge graph and used as constraints for the generation model to forcefully inject generation prompts, thereby ensuring the consistency of asset visual features across multiple shots.

9. The comic book generation system according to claim 8, characterized in that: The agent for distribution is configured as follows: If the user inputs a single sentence of text or a single image, the intelligent agent will be automatically invoked to expand or understand the material and complete the entire workshop intelligent agent process. If the user input is a video clip, then reverse reasoning starts from the storyboard agent or resource agent to complete the script or material information in the forward direction; If the user input is a long text or a script document, all workshop agents are scheduled in the standard order.

Images