Short video generation method and apparatus

Abstract

The embodiment of the specification provides a short video generation method and device. The method comprises: acquiring a demand description input by a user for a target short video to be generated; then recalling product knowledge related to a target insurance product from a knowledge base based on the demand description, the product knowledge comprising an insured user portrait, an accident scene and a policyholder-beneficiary relationship; taking the demand description and the product knowledge as an input context of an intelligent agent system to guide the intelligent agent system to generate the target short video, wherein a role image in the target short video is generated by the intelligent agent system with reference to the insured user portrait, a plot in the target short video is generated by the intelligent agent system with reference to the accident scene, and a character relationship and a dialogue in the target short video are generated by the intelligent agent system with reference to the policyholder-beneficiary relationship. In this way, the generated short video has precise crowd matching degree and real scene substitution, and the quality of the generated short video is enhanced.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of artificial intelligence (AI) technology, and more particularly to a short video generation method and apparatus. Background Technology

[0002] With the rise of short videos, they have gradually become one of the most popular and effective forms of advertising content. In marketing, short videos can attract user attention, convey information, and achieve promotional goals. For example, in the financial and insurance industry, an increasing number of insurance agents are using short videos to introduce insurance products, interpret insurance terms, popularize insurance types, and promote insurance products.

[0003] Traditional short video production requires a lot of manual production work such as shooting and editing. Insurance agents often need to invest a lot of manpower, resources and time to create high-quality short videos during the video production and audio recording process. This is not only time-consuming and costly, but also clearly cannot meet the actual needs of the financial insurance industry, which requires rapid response to market changes and has a large number of products.

[0004] While using AI to generate short videos can improve the efficiency of short video production, existing AI-generated short videos have many problems, especially in the financial and insurance industry, where AI-generated short videos suffer from issues such as severe content homogenization and a lack of storytelling. Summary of the Invention

[0005] This specification describes one or more embodiments of a short video generation method and apparatus that ensures the generated short videos have accurate audience matching and a realistic sense of scene immersion, thereby enhancing the quality of the generated short videos.

[0006] According to the first aspect, a short video generation method is provided, including:

[0007] Obtain the user's input description of their requirements for the target short video to be generated;

[0008] Based on the aforementioned requirements description, product knowledge related to the target insurance product is retrieved from the knowledge base. This product knowledge includes the insured user profile, the accident scenario, and the relationship between the policyholder and the insured.

[0009] The demand description and product knowledge are used as the input context of the intelligent agent system to guide the system in generating a target short video. The character images in the target short video are generated by the intelligent agent system with reference to the insured user profile, the plot of the target short video is generated by the intelligent agent system with reference to the accident scenario, and the relationships and dialogues between the characters in the target short video are generated by the intelligent agent system with reference to the relationship between the insured and the policyholder.

[0010] According to the second aspect, a short video generation device is provided, comprising:

[0011] The acquisition module is used to acquire the user's input description of the target short video to be generated;

[0012] The recall module is used to recall product knowledge related to the target insurance product from the knowledge base based on the requirement description. The product knowledge includes the insured user profile, the accident scenario, and the relationship between the insured and the policyholder.

[0013] The generation module is used to take the requirement description and the product knowledge as the input context of the intelligent agent system, and guide the intelligent agent system to generate a target short video. The character images in the target short video are generated by the intelligent agent system with reference to the insured user profile, the plot of the target short video is generated by the intelligent agent system with reference to the accident scenario, and the relationships and dialogues between the characters in the target short video are generated by the intelligent agent system with reference to the relationship between the insured and the policyholder.

[0014] According to a third aspect, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the method described in the first aspect.

[0015] According to a fourth aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described in the first aspect.

[0016] According to a fifth aspect, a computing device is provided, including a memory and a processor, wherein executable code is stored in the memory, and the processor, when executing the executable code, implements the method of the first aspect.

[0017] In the embodiments of this specification, product knowledge related to the target insurance product is retrieved from the knowledge base based on the user's input requirement description, including the insured user profile, the accident scenario, and the relationship between the insured and the policyholder. This knowledge is then injected into the input context of the intelligent agent system, guiding the system to generate a character image that matches the insured user profile, a plot that revolves around the accident scenario, and character relationships and dialogues based on the relationship between the insured and the policyholder during the short video generation process. This ensures that the generated short video has accurate audience matching, a realistic sense of scene immersion, and enhances the quality of the generated short video. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This diagram illustrates an application scenario of the short video generation method provided in the embodiments of this specification.

[0020] Figure 2 A system architecture diagram is shown that can implement the short video generation method provided in the embodiments of this specification;

[0021] Figure 3 A flowchart illustrating a short video generation method according to one embodiment is shown;

[0022] Figure 4 A schematic diagram of a short video generation apparatus according to one embodiment is shown. Detailed Implementation

[0023] To facilitate understanding of the solutions provided in the embodiments of this specification, a brief introduction will first be given to the technical terms that may be involved in the embodiments of this specification.

[0024] Intelligent Agent: Refers to an intelligent software entity built on a large model, possessing autonomous perception, planning, decision-making, and execution capabilities. In this solution, it specifically refers to a vertical domain intelligent agent that integrates insurance domain knowledge, delivery strategies, and script creation capabilities.

[0025] AI Generated Content (AIGC): This solution encompasses AI-generated video character designs, AI-generated storyboards, AI-generated video footage, and post-production editing elements.

[0026] OpenClaw SKILL: A skill unit that standardizes, encapsulates, and consolidates agent capabilities, supporting invocation and reuse.

[0027] Content placement: refers to placing short videos on short video platforms.

[0028] Accident scenarios: Based on historical insurance claims data after anonymization, the types of real risk events and their occurrence situations are extracted (such as elderly people slipping in the bathroom, being scalded while cooking, falling down stairs, etc.).

[0029] Policyholder-Insured Relationship: The family relationship between the policyholder and the insured in the insurance contract, such as the policyholder, spouse, parents, children, etc.

[0030] High-conversion creative characteristics: Short videos that have been proven to have high click-through rates and conversion rates (such as completion rate and order rate) in historical campaigns.

[0031] Storyboard: In short video production, independent shot units are divided according to time sequence. Each storyboard includes a scene description, dialogue, duration, shot size, camera movement, etc.

[0032] Senior Comprehensive Accident Insurance: This is short for Senior Comprehensive Accident Insurance, which mainly covers the risks of death, disability, and medical expenses caused by accidents to seniors.

[0033] Return on Investment (ROI): A core metric for measuring the effectiveness of advertising campaigns.

[0034] Digital humans: Virtual characters generated through computer graphics and AI technology, used to replace real people in creating video content.

[0035] Poster-style materials: Static or simple animated short video materials with text titles and product selling point lists as the main visual elements.

[0036] To meet the financial and insurance industry's demand for efficient short video generation, most related technical solutions employ AI-generated methods to create insurance short videos. For example, one such solution proposes a digital human-based narration video generation method. Specifically, the user inputs text or selects a template, and the system uses an AI-generated digital human avatar to narrate the content, combined with a text title and a simple background, to generate the short video content.

[0037] This solution has the following drawbacks:

[0038] 1. The content is highly homogenized and lacks narrative. Digital human videos are essentially "virtual avatars reading text," with a monotonous presentation format, no scene changes, no character interaction, and no emotional progression. In the context of news feed ads, users have extremely low tolerance for this kind of "machine-read" content and usually swipe away within 3 seconds.

[0039] 2. Lack of industry knowledge. General digital human platforms do not understand the characteristics of insurance products, and the copy they generate is often a simple combination of "product selling points + promotional language", which fails to address the real pain points of users' insurance needs.

[0040] 3. The digital persona lacks realism and trustworthiness. Although digital humans are constantly advancing in technology, in the marketing of financial products such as insurance, which require high trust, the digital persona naturally lacks the credibility and approachability of a real person. Users have an instinctive distrust of insurance products recommended by digital humans.

[0041] 4. Inability to achieve multi-character dialogue. Digital human solutions typically feature a single character, making it impossible to present dialogue and interaction between characters, thus limiting the depth of emotional expression and the storytelling.

[0042] 5. Lack of data-driven creative optimization capabilities. The digital human solution cannot automatically optimize creative content and style based on user profiles, historical campaign data, and incident scenarios; the quality of the creatives relies entirely on the quality of manually entered copy.

[0043] Furthermore, in one related technical solution, a general AIGC video generation scheme is adopted. Specifically, the user inputs a text description, and the AI ​​model directly generates video content. Alternatively, the user inputs an image and a text description, and the AI ​​model generates video content.

[0044] This solution has the following drawbacks:

[0045] 1. The generated content is uncontrollable and cannot meet insurance compliance requirements. Insurance marketing materials have strict compliance red lines (such as prohibiting the use of absolute terms, prohibiting promises of returns, and requiring the full presentation of product names, etc.). The output of general AIGC video generation models is random and unpredictable, and it cannot be guaranteed that every frame and every line generated will meet the regulatory requirements of the insurance industry.

[0046] 2. Lack of structured output capabilities. General video generation solutions output "end-to-end" video files, which cannot provide structured control at the scene level. However, insurance advertising materials require precise control over the content, dialogue, duration, and pacing of each scene to achieve optimal conversion results.

[0047] 3. Inconsistency in character portrayal cannot be guaranteed. In short videos that require multiple shots and scenes, the general solution cannot guarantee the consistency of the same character's image in different shots, resulting in a strong sense of disjointedness in the visuals.

[0048] 4. High production costs and unstable efficiency. The quality of currently generated AIGC videos fluctuates greatly, and the probability of generating satisfactory content on the first attempt is low. Multiple trials and selections are required, resulting in an unstable output efficiency of usable material.

[0049] 5. Lack of knowledge in the insurance field. The general solution fails to understand the differences between "accident insurance for the elderly" and "critical illness insurance for children" in terms of target groups, pain points, and compliance requirements, resulting in materials that lack industry relevance.

[0050] Another related technical solution proposes a template-based material generation method. Specifically, it provides preset video templates, allowing users to quickly generate video materials by replacing elements such as text and images.

[0051] This solution has the following drawbacks:

[0052] 1. Excessive template-based advertising and homogenized content. The same template is used by numerous advertisers, leading to highly similar content on the platform. This causes user fatigue, and the platform's algorithm also lowers the recommendation weight of similar content.

[0053] 2. Lack of industry customization. Generic templates do not consider the characteristics of the insurance industry and fail to present the sense of scenario, trust, and emotional resonance required by insurance products.

[0054] 3. Unable to automatically generate creative ideas. Template tools only solve the efficiency problem of the "production" stage, but the "creative" stage still needs to be completed manually, and does not fundamentally solve the core bottleneck of material production.

[0055] In summary, the short video generation solutions provided by the relevant technologies have various defects and shortcomings.

[0056] In view of this, embodiments of this specification propose a short video generation method. This method retrieves product knowledge related to the target insurance product from a knowledge base based on the user's input description of their needs. This knowledge includes the insured user profile, the accident scenario, and the relationship between the insured and the policyholder. This knowledge is then injected into the input context of an intelligent agent system, guiding the system to generate character images that match the insured user profile, a storyline revolving around the accident scenario, and character relationships and dialogues based on the insured and policyholder relationship. This ensures that the generated short video has accurate audience matching, a realistic sense of scene immersion, and enhances the quality of the generated short video.

[0057] The specific implementation of the short video generation method proposed in this specification is described in detail below with reference to the accompanying drawings and embodiments.

[0058] Figure 1 This diagram illustrates an application scenario of the short video generation method provided in the embodiments of this specification. For example... Figure 1As shown, when a user needs to generate a short video, they can open the short video generation client on client device 100 and input a requirement description. The short video generation client generates a short video generation request based on the user's input requirement description and sends the request to server device 200. The short video generation service deployed on server device 200 responds to the request, parses it to obtain the requirement description, and then executes the short video generation method provided in this embodiment to generate a short video that meets the user's requirements.

[0059] After generating the short video, the server device 200 sends the short video to the client device 100 so that the user can view the generated short video content on the client device 100.

[0060] Client device 100 may include, but is not limited to, devices such as smartphones, tablets, laptops, and personal computers.

[0061] The server-side device 200 can be implemented through a single server or a server cluster consisting of multiple servers. The server can be a local server or a cloud server.

[0062] The server device 200 is equipped with a short video generation system that can implement the short video generation method provided in this disclosure and can provide short video generation services to external users.

[0063] The short video generation service can generate short videos to introduce insurance products, helping insurance agents better present these products to clients. It can also generate short videos for marketing insurance products, helping agents promote them more effectively. The type of short video generated depends on the input requirements; this specification does not limit the specific type of short video generated by the service.

[0064] Figure 2 A system architecture diagram is shown that can implement the short video generation method provided in the embodiments of this specification. Figure 2 As shown, the system comprises a data and knowledge layer, an agent layer, an AIGC generation and synthesis layer, and a delivery and feedback layer. Among these,

[0065] The data and knowledge layer is used to aggregate and manage all the data sources and domain knowledge required by the system.

[0066] Optionally, the data and knowledge layer includes an insured user profile database, an accident scenario knowledge graph database, an insured-policy relationship knowledge database, a high-conversion material feature database, an insurance product knowledge database, and a compliance rule database.

[0067] For example, the insured user profile database is used to store insured user profiles, and each insured user profile includes multi-dimensional feature data such as age, occupation, region, and family structure of the target insured population.

[0068] The accident scenario knowledge graph is used to store accident scenario knowledge. This knowledge graph is built based on anonymized historical claims data and includes structured knowledge such as accident type (e.g., falls, fractures, burns, traffic accidents), accident location (e.g., bathroom, kitchen, stairs, outdoors), typical cost ranges (e.g., hip replacement surgery costs, length of hospital stay), high-risk population characteristics, policyholder profiles, and policyholder-insured relationship nodes. In other words, the accident scenario knowledge graph includes nodes for accident type, accident location, typical cost range, high-risk population characteristics, policyholder profiles, and policyholder-insured relationship.

[0069] The policyholder-insured relationship knowledge base is used to store the policyholder-insured relationship, including different relationship types such as the policyholder, the policyholder insured for their parents, the policyholder insured for their spouse, and the policyholder insured for their children, as well as the typical psychological motivations and communication scenarios under each relationship type.

[0070] The high-conversion creative feature library is used to store script structure features extracted from historical high ROI campaign creatives. These script structure features include structured experiences such as pain point triggering methods, dialogue rhythm, scene type distribution, and character relationship patterns.

[0071] The insurance product knowledge base is used to store knowledge about multiple insurance products. Each insurance product knowledge includes structured product information such as product terms, coverage, premium information, compliance constraints, and product selling points.

[0072] The compliance rules base is used to store compliance knowledge for insurance marketing, such as compliance red lines for insurance marketing (prohibited words, required statements, product name specifications, etc.).

[0073] The Agent layer is the core layer for short video generation. It includes at least one Agent. Through Prompt Engineering, data and knowledge layer content are systematically injected into the Agent's context, guiding the Agent to generate content based on the injected insurance domain knowledge.

[0074] Optionally, the Agent layer includes multiple Agents, which can be functionally categorized as Creative Agent, Director Agent, and Editor Agent. The Creative Agent is responsible for understanding user input (such as user-inputted needs descriptions), accessing multi-source information from the data and knowledge layers, and completing tasks such as user profiling analysis, pain point selection, scene setting, character design (including character image descriptions, also known as character DNA descriptions), dialogue script creation, and product benefit presentation, ultimately outputting a short video script.

[0075] The Director Agent is responsible for transforming the short video scripts output by the Creative Agent into visual solutions, including character AIGC image generation (based on character image description), storyboard image generation (based on storyboard description), and video clip generation, outputting multiple video clips.

[0076] The editor, Agent, is responsible for post-production, including adding watermarks, generating and overlaying subtitles, matching sound effects, selecting and mixing background music, generating and overlaying subtitles, and finally outputting a complete short video (i.e., the target short video).

[0077] The AIGC generation and synthesis layer is responsible for transforming the Agent's output into actual visual and audio content.

[0078] Optionally, the AIGC generation and compositing layer includes a character image generation engine, a storyboard image generation engine, a speech synthesis engine, and a video synthesis engine.

[0079] Among them, the character image generation engine is used to generate character image images with set parameters (such as 9:16 ratio, 1K resolution) based on the character image description and by calling the text-based image model (such as StableDiffusion, DALL·E, Midjourney, etc.).

[0080] In practical applications, the Director Agent can call the character image generation engine to generate character images based on the character image descriptions in the short video script output by the Creative Agent.

[0081] The storyboard generation engine is used to generate corresponding storyboard frames based on storyboard descriptions by calling text-to-image / image-to-video models.

[0082] In practical applications, the Director Agent can call the storyboard generation engine to generate storyboards based on the storyboard descriptions in the short video script output by the Creative Agent.

[0083] The speech synthesis engine is used to generate the voice of the corresponding character based on the script dialogue, and supports emotional tone adjustment.

[0084] The video compositing engine is used to combine elements such as storyboard footage, audio, background music, sound effects, subtitles, and watermarks into the final video material for deployment.

[0085] The delivery and feedback layer is used for channel delivery adaptation and delivery performance feedback.

[0086] Channel adaptation means automatically adjusting the parameters of short videos (such as video format, duration, resolution, etc.) according to the requirements of the advertising channel. For example, if the first short video platform requires the short video to use the first parameter combination, then the parameters of the generated short video will be adjusted to the first parameter combination through channel adaptation; if the second short video platform requires the short video to use the second parameter combination, then the parameters of the generated short video will be adjusted to the second parameter combination through channel adaptation.

[0087] The campaign performance recovery is used to collect feedback data on the short video campaign (such as exposure, click-through rate, conversion rate, completion rate, etc.). Then, based on the feedback data, high-conversion creatives are identified. High-conversion creative features with script structure characteristics are extracted from the high-conversion creatives. These high-conversion creative features are then added to the high-conversion creative feature library in the data and knowledge layer, thus continuously optimizing the high-conversion creative feature library.

[0088] The following describes the detailed implementation of the short video generation method provided in the embodiments of this specification.

[0089] Figure 3 A flowchart illustrating a short video generation method according to one embodiment is shown. This method can... Figure 1 The server device 200 shown implements this, and the server device 200 can be any device, equipment, platform, or cluster of devices with computing and processing capabilities.

[0090] In step S301, the user's input description of their requirements for the target short video to be generated is obtained.

[0091] The short video generation system can parse the received short video generation request to obtain the user's requirement description for the target short video to be generated.

[0092] For example, if an insurance agent wants to generate a short marketing video for an insurance product, they can input a requirement description through a short video generation client on their client device. The short video generation client generates a short video generation request based on the user's input requirement description and sends the request to the server device. The short video generation service deployed on the server device 200 responds to the request and parses it to obtain the requirement description.

[0093] User input can be a simple one-sentence request, including only the target product information. For example, a user might input: "Help me generate a short marketing video for comprehensive accident insurance for seniors." User input can also include more information beyond the target product, such as the video's theme, plot, characters, setting, style, and duration. For example, a user might input: "Make a 30-second car insurance video targeting car owners aged 25-45, highlighting a 3-minute online quote, an average saving of 20% compared to offline purchases, a fast-paced opening emphasizing cost-saving anxiety, and ending with a QR code and limited-time offers."

[0094] This disclosure does not limit the specific form of the user-inputted requirement description. Users can input a concise requirement description or a more detailed requirement description according to their actual needs.

[0095] In step S302, product knowledge related to the target insurance product is retrieved from the knowledge base based on the demand description. This product knowledge includes the insured user profile, the accident scenario, and the relationship between the insured and the policyholder.

[0096] After obtaining the demand description, the short video generation system will retrieve relevant product knowledge of the target insurance product from the knowledge base based on the demand description. The retrieved product knowledge includes at least the insured user profile, the accident scenario, and the relationship between the policyholder and the insured.

[0097] For example, the short video generation system uses the requirement description as a query to retrieve knowledge related to the current short video generation task from the data and knowledge layer. This relevant knowledge may include insured user profiles, accident scenarios, and the relationship between the insured and the policyholder.

[0098] For example, the short video generation system first uses the requirement description as a query statement to recall accident scenarios related to this short video generation task from the accident scenario knowledge graph library. The recalled accident scenario knowledge includes at least the accident type, accident occurrence scenario, typical cost range, high-risk population characteristics, insured user profile, and the relationship between the insured and the policyholder.

[0099] Optionally, the policyholder profile and policyholder-insured relationship retrieved in the event of an accident are both represented by IDs. Then, the policyholder profile ID is used to retrieve detailed policyholder profile information from the policyholder profile database, including multi-dimensional characteristics such as age, occupation, region, and family structure. The policyholder-insured relationship ID is used to retrieve the policyholder-insured relationship from the policyholder-insured relationship knowledge base, including different relationship types such as the policyholder themselves, policyholders insured for their parents, policyholders insured for their spouse, and policyholders insured for their children. This also includes typical psychological motivations and communication scenarios for each relationship type.

[0100] This allows for the retrieval of product knowledge related to the target insurance product from the knowledge base using the requirements description.

[0101] In step S303, the requirement description and product knowledge are used as the input context for the intelligent agent system to guide the intelligent agent system in generating the target short video.

[0102] After successfully retrieving product knowledge related to the target insurance product, the short video generation system uses the user's input description of needs and product knowledge as the input context for the intelligent agent system, guiding the intelligent agent system to generate the target short video.

[0103] In other words, when generating short videos, the Agent system does not rely solely on the user's input description of needs, but systematically integrates data sources from at least three dimensions: the insured user profile, the accident scenario, and the relationship between the insured and the policyholder.

[0104] In this way, the Agent system can generate character images by referring to the profile of the insured user, generate the plot of the target short video by referring to the accident scenario, and generate the character relationships and dialogues in the target short video by referring to the relationship between the insured and the policyholder.

[0105] In other words, the insured user profile determines the age, occupation, and language style of the characters in the dialogue. The accident scenario determines the type of accident, the location of the accident, and the costs in the story. The relationship between the insured and the policyholder determines the relationship between the characters (e.g., spouses, sisters, father and daughter) and the emotional tone of the dialogue.

[0106] This disclosure embodiment ensures that the generated short video materials have: accurate audience matching, realistic scene immersion, and appropriate emotional tension by injecting knowledge retrieved from three-dimensional data sources into the creation context of the Agent system, thereby enhancing the quality of the generated short videos.

[0107] In one example, to further improve the quality of the generated short video, the Agent system of this disclosure uses a collaborative approach of multiple Agents to generate the short video.

[0108] For example, the short video generation process can be divided into three stages: creative, directing, and editing, each executed by a corresponding agent. The creative agent focuses on creating the short video script, determining the overall creative framework. The directing agent is responsible for bringing the creative agent's ideas to life, generating video clips, and ensuring the quality of those clips. The editing agent focuses on editing multiple video clips to improve the overall smoothness of the video. This division of labor allows each stage of video generation to receive professional processing, improving the accuracy of each stage and enhancing the match between the final generated video and user needs. This results in a better integration of visuals, audio, and content expression, further improving the overall quality of the final short video.

[0109] In actual execution, the user's input of requirements and product knowledge serves as input for the creative agent, outputting a structured JSON-formatted short video script. Optionally, this script includes scene numbers, dialogue, duration, shot size, camera movement, sound effects, character descriptions, scene descriptions, and a plot summary. The character descriptions are generated by the creative agent with reference to the insured user profile; the plot summary is generated by the creative agent with reference to the accident scenario; and the character relationships and dialogue are generated by the creative agent with reference to the relationship between the policyholder and the insured.

[0110] By introducing real-life accident scenarios, each script unfolds a complete story around a specific, authentic accident situation, increasing the authenticity of the short video content. Furthermore, character designs are generated based on insured user profiles, ensuring accurate audience matching; the characters' appearances are similar to the matched demographics, enhancing user immersion. In addition, character relationships and dialogues are generated based on the relationship between the insured and the policyholder, enhancing emotional interaction between characters and further building engagement.

[0111] After the Creative Agent generates the short video script, the short video script is used as input to the Director Agent. Based on the character image description and storyboard description in the short video script, the Director Agent calls the AIGC model to generate character image diagrams and storyboard images, and then synthesizes the storyboard images into multiple video clips for output.

[0112] After the Creative Agent generates a short video script and the Director Agent generates multiple video clips, the Editor Agent is invoked to edit the video footage output by the Director Agent. Specifically, the short video script and multiple video clips are used as input to the Editor Agent, and the target short video is output.

[0113] The editor (Agent) is responsible for post-production of the video footage output by the director (Agent), such as adding watermarks, volume down, sound effects, background music, and subtitles, and finally outputting a coherent and smooth complete short video with audio, i.e., the target short video.

[0114] This three-stage separation short video generation architecture allows for independent inspection and correction of the output at each stage, ensuring precise control over each phase. Furthermore, different stages can utilize suitable AIGC models, enabling flexible orchestration and optimization of the overall process.

[0115] Given the unique nature of the insurance industry, various legal and industry compliance requirements exist. This disclosed embodiment also includes an insurance product knowledge base and a compliance rule base in the data and knowledge layer. During product knowledge retrieval, compliance constraints on the target insurance product will be retrieved from the insurance product knowledge base, and compliance constraints on insurance marketing will be retrieved from the compliance rule base.

[0116] The compliance constraints of the target insurance product and insurance marketing are also injected into the input context of the creative agent. This allows the creative agent to generate short video scripts under the constraints of these two constraints. In this way, compliance constraints are embedded into the creative agent's generation logic as hard constraints, ensuring that the short video scripts generated by the creative agent are generally compliant, achieving pre-emptive compliance injection (i.e., injecting compliance constraints during the short video script generation stage).

[0117] Understandably, the insurance product knowledge base stores the compliance constraints for each insurance product, while the compliance constraints for the target insurance product in the recall are specific to that product and can be understood as vertical constraints. The compliance rule base stores compliance constraints for the insurance marketing field, while the compliance constraints for the recalled insurance marketing apply to all insurance products. In other words, all insurance products must comply with these constraints during the marketing process, which can be understood as horizontal constraints.

[0118] In one example, to further ensure that the short video scripts generated by the creative agent are compliant, the agent system in this embodiment is also configured with a script compliance review agent. The script compliance review agent is used to review whether the short video scripts generated by the creative agent are compliant, that is, whether they comply with the compliance constraints of the target insurance product and the compliance constraints of insurance marketing. If they comply, the compliance review is passed; if they do not comply, the compliance review is failed.

[0119] For example, after the Creative Agent generates a short video script, the system will call the Script Compliance Review Agent to conduct a compliance review of the short video script. For example, the Script Compliance Review Agent will check each item:

[0120] Did it use absolute terms such as "first-of-its-kind," "exclusive," "highest," or "only"?

[0121] Does it make any uncertain promises or guarantees regarding future returns?

[0122] Does it denigrate or compare with other insurance products or companies?

[0123] Does the product name need to be presented in full (e.g., the full name of the insurance product must be "******")?

[0124] Whether specific terms comply with the substitution rules (e.g., "premium" is replaced with "·premium", "fracture" is replaced with "·fracture", etc.).

[0125] If any non-compliant content is detected in the short video script, the script compliance review agent will output "review failed" and generate modification suggestions (e.g., which content is non-compliant and how to modify it). These suggestions are then injected into the input context of the creative agent, guiding it to regenerate the short video script. The script compliance review agent is then invoked again to conduct a compliance review of the short video script until it passes, ensuring the final generated short video script is compliant.

[0126] This ensures that the generated short video scripts are compliant; for example, it ensures that every line of dialogue and every product description complies with the regulatory requirements of the insurance industry.

[0127] In another example, the output of the director agent also includes character images and storyboards; that is, the director agent outputs character images, storyboards, and multiple video clips. To ensure that the visual content generated by the director agent is compliant, such as the character images and storyboards, the agent system in this embodiment is also configured with a visual compliance review agent. The visual compliance review agent is used to review whether the character images and storyboards generated by the director agent are compliant, that is, whether they comply with the compliance constraints of the target insurance product and the compliance constraints of insurance marketing. If they comply, the compliance review passes; if they do not comply, the compliance review fails.

[0128] For example, after the Director Agent generates the character image and storyboard, the system calls the Visual Compliance Review Agent to review them for compliance. If any violations are detected (e.g., the target insurance product is two-wheeled vehicle insurance, but the character is not wearing a helmet), the Visual Compliance Review Agent will output "Review Failed" and generate adjustment suggestions (e.g., which content is non-compliant and how to adjust it). These suggestions are then injected into the Director Agent's input context, guiding it to regenerate the character image and / or storyboard (if only the character image is non-compliant, only the character image is regenerated; if the storyboard is non-compliant, only the storyboard is regenerated; if both the character image and storyboard are non-compliant, both must be regenerated). The system then continues to call the Visual Compliance Review Agent to review the regenerated character image and / or storyboard until the compliance review is passed, ensuring that the final generated visual content is compliant.

[0129] This ensures that the visual content output by the director agent is also compliant, meaning that the output visual content meets the regulatory requirements of the insurance industry.

[0130] Optionally, the Agent system in this embodiment can also be configured with a script compliance review Agent and a visual compliance review Agent simultaneously, thus achieving dual compliance review. That is, at the text level, the script compliance review Agent performs semantic-level review on the short video script generated by the creative agent; at the visual level, the visual compliance review Agent performs media content compliance review on the visual content (such as character images and storyboards) generated by the director agent, ensuring the compliance of the final generated short video content.

[0131] Thus, the Agent system comprises five agents: Creative Agent, Director Agent, Editor Agent, Script Compliance Review Agent, and Visual Compliance Review Agent. Among them, the Creative Agent, Director Agent, Editor Agent, and Script Compliance Review Agent are all built based on a large language model, while the Visual Compliance Review Agent is built based on a large visual model.

[0132] In another example, the data and knowledge layer also includes a high-conversion material feature library, which stores multiple script structure features, each extracted from high-conversion short video materials. Each script structure feature includes structured experiences such as pain point triggering methods, dialogue rhythm, scene type distribution, and character relationship patterns.

[0133] Understandably, high-converting short video content refers to short video content that has been proven to have high conversion rates through experience.

[0134] The recalled product knowledge also includes high-conversion creative features (i.e., script structure features) relevant to this generation, retrieved from a high-conversion creative feature library. The input context of the Creative Agent also incorporates these high-conversion creative features. During script generation, the Creative Agent references these features, such as the timing of dialogue pacing and the introduction of the target insurance product. In this way, the Creative Agent references high-conversion creative features during the generation process, increasing the potential of the generated short video to become a viral hit.

[0135] In another example, the short video generation method provided in this disclosure further includes: obtaining feedback data of the short video after it has been deployed; then identifying high-conversion short video materials based on the feedback data, and extracting script structure features from the high-conversion short video materials; and finally updating the high-conversion material feature library based on the extracted script structure features.

[0136] For example, after the final short video is deployed to a short video platform, the system collects feedback data such as impressions, click-through rate (CTR), conversion rate, and completion rate. Then, it evaluates the short video's score across multiple dimensions. For instance, the scores for these four dimensions are Score_1, Score_2, Score_3, and Score_4. These scores are then weighted and summed to obtain the final evaluation score, Score_f. The weighting of each dimension can be pre-set. For example, if the weighting of impressions is 0.1, CTR is 0.2, conversion rate is 0.4, and completion rate is 0.3, then the final evaluation score Score_f = 0.1 * Score_1 + 0.2 * Score_2 + 0.4 * Score_3 + 0.3 * Score_4. Short videos with an evaluation score greater than a preset threshold (e.g., 0.8) are identified as high-conversion content.

[0137] Script structure features are extracted from the identified high-conversion content. For example, a multimodal large model is used to extract script structure features from high-conversion content, such as dialogue rhythm, pain point triggering methods, and product introduction timing.

[0138] Finally, the extracted script structure features of high-conversion materials are added to the high-conversion material library to enrich the library.

[0139] In this way, the feature library of high-conversion materials is continuously updated, and the creative strategies of creative agents are constantly evolved, forming a closed-loop evolution of "generation → distribution → recycling → optimization → generation".

[0140] In some other examples, embodiments of this disclosure may also encapsulate the capabilities of the Agent system described above into a SKILL (e.g., an OpenClaw SKILL). This allows the SKILL to be reused by different business scenarios and callers without repeated development, and the SKILL can be used as a node in a larger workflow in combination with other SKILLs.

[0141] Optionally, the encapsulated SKILL has a standardized interface and standardized output. The standardized interface is configured to receive input in a uniform format (e.g., insurance type, quantity of materials, material format, and distribution channel). The standardized output is to output a structured JSON short video script and a finished short video in a uniform format.

[0142] In summary, the technical concepts of the various embodiments in this specification integrate at least three core data sources from the insurance industry—insured user profiles, accident scenarios, and the relationship between the insured and the policyholder—into the creative context of the Agent system, enabling the automatic generation of short video scripts with industry depth and user precision. This represents the first successful leap from "general text summarization → video" to "multi-dimensional industry data → precise script → video." Secondly, the end-to-end short video generation process is decomposed into three independent but sequential stages: creative (script layer), director (visual layer), and editor (post-production layer), each handled by a dedicated Agent. This decoupled architecture allows for independent verification and optimization of the output of each stage, ensuring precise controllability at each stage and enabling the selection of appropriate AIGC models for different stages. The overall process can be flexibly arranged and optimized. Thirdly, by injecting compliance constraints into the context of the creative Agent, and by conducting compliance reviews of the short video scripts generated by the creative Agent and the visual materials generated by the director Agent, the compliance of the generated content is guaranteed. Finally, by collecting actual performance data of the campaign materials (such as impressions, click-through rate, conversion rate, and completion rate), high-conversion materials are identified, their characteristics are extracted, and the high-conversion material feature library is updated. This allows the Agent's creation strategy to be continuously optimized based on real campaign data, forming a closed-loop evolution of "generation → deployment → recycling → optimization → generation".

[0143] According to another embodiment, a short video generation apparatus is provided. Figure 4 The diagram illustrates the structure of a short video generation apparatus according to one embodiment. This apparatus can be deployed in any device, platform, or device cluster with data storage, computing, and processing capabilities. Figure 4 As shown, the device 400 includes:

[0144] The acquisition module 401 is used to acquire the user's input description of the requirements for the target short video to be generated;

[0145] The recall module 402 is used to recall product knowledge related to the target insurance product from the knowledge base based on the requirement description. The product knowledge includes the profile of the insured user, the accident scenario, and the relationship between the insured and the policyholder.

[0146] The generation module 403 is used to take the requirement description and the product knowledge as the input context of the intelligent agent system to guide the intelligent agent system to generate a target short video. The character images in the target short video are generated by the intelligent agent system with reference to the insured user profile, the plot of the target short video is generated by the intelligent agent system with reference to the accident scenario, and the relationships and dialogues between the characters in the target short video are generated by the intelligent agent system with reference to the relationship between the insured and the policyholder.

[0147] In one embodiment, the intelligent agent system includes a creative agent, a director agent, and an editor agent;

[0148] The generation module 403 is specifically used for:

[0149] The input context is used as the input of the creative agent to output a short video script. The character descriptions in the short video script are generated by the creative agent with reference to the insured user profile. The plot summary in the short video script is generated by the creative agent with reference to the accident scene. The character relationships and lines in the short video script are generated by the creative agent with reference to the relationship between the insured and the policyholder.

[0150] The short video script is used as input to the director's intelligent agent to output multiple video clips;

[0151] The short video script and the multiple video clips are used as input to the editor agent to output the target short video.

[0152] According to one embodiment, the product knowledge also includes compliance constraints on the target insurance product and compliance constraints on insurance marketing;

[0153] The creative agent generates the short video script under the constraints of compliance with the target insurance product and the insurance marketing.

[0154] According to one embodiment, the intelligent agent system further includes a script compliance review intelligent agent;

[0155] The short video generation device further includes: a compliance review module 404, which is used for:

[0156] The script compliance review agent is invoked to conduct a compliance review of the short video script. If the review fails, modification suggestions are generated.

[0157] The proposed modifications are injected into the input context of the creative agent, guiding the creative agent to regenerate the short video script.

[0158] According to one embodiment, the output of the director's intelligent agent also includes character images and storyboards;

[0159] The intelligent agent system also includes a visual compliance review intelligent agent;

[0160] The compliance review module 404 is also used for:

[0161] The visual compliance review agent is invoked to conduct a compliance review of the character image and storyboard. If the review fails, adjustment suggestions are generated.

[0162] The adjustment suggestions are injected into the input context of the director agent, guiding the director agent to regenerate the character image and storyboard.

[0163] According to another embodiment, the product knowledge also includes high-conversion material features, which are script structure features of short videos extracted from short video materials with high conversion rates related to the target insurance product;

[0164] The dialogue rhythm and the timing of introducing the target insurance product in the short video script are generated by the creative agent with reference to the characteristics of the high-conversion materials.

[0165] According to one implementation, the knowledge base includes a high-conversion material feature library, which includes multiple script structure features extracted from high-conversion short video materials;

[0166] The short video generation device further includes a feedback update module 405, which is used for:

[0167] Obtain feedback data on the short videos after they are deployed;

[0168] High-conversion short video materials are identified based on the feedback data;

[0169] Script structure features were extracted from the high-conversion short video material;

[0170] The high-conversion material feature library is updated based on the extracted script structure features.

[0171] For specific implementation examples of each unit in the above device, please refer to the previous examples. Figure 3 The above devices enable the generation of short videos to achieve accurate audience matching, realistic scene immersion, and enhanced video quality.

[0172] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform a combination Figure 3 The method described.

[0173] According to yet another embodiment, a computer program product is also provided, including a computer program / instructions that, when executed by a processor, implement the foregoing combinations. Figure 3 The described method and steps.

[0174] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements a combination... Figure 3 The method described.

[0175] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0176] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating short videos, comprising: Obtain the user's input description of their requirements for the target short video to be generated; Based on the aforementioned requirements description, product knowledge related to the target insurance product is retrieved from the knowledge base. This product knowledge includes the insured user profile, the accident scenario, and the relationship between the policyholder and the insured. The demand description and product knowledge are used as the input context of the intelligent agent system to guide the system in generating a target short video. The character images in the target short video are generated by the intelligent agent system with reference to the insured user profile, the plot of the target short video is generated by the intelligent agent system with reference to the accident scenario, and the relationships and dialogues between the characters in the target short video are generated by the intelligent agent system with reference to the relationship between the insured and the policyholder.

2. The method according to claim 1, wherein, The intelligent agent system includes a creative agent, a director agent, and an editor agent; The step of using the demand description and product knowledge as input context for the intelligent agent system to guide the system in generating the target short video includes: The input context is used as the input of the creative agent to output a short video script. The character descriptions in the short video script are generated by the creative agent with reference to the insured user profile. The plot summary in the short video script is generated by the creative agent with reference to the accident scene. The character relationships and lines in the short video script are generated by the creative agent with reference to the relationship between the insured and the policyholder. The short video script is used as input to the director's intelligent agent to output multiple video clips; The short video script and the multiple video clips are used as input to the editor agent to output the target short video.

3. The method according to claim 2, wherein, The product knowledge also includes compliance constraints on the target insurance product and compliance constraints on insurance marketing; The creative agent generates the short video script under the constraints of compliance with the target insurance product and the insurance marketing.

4. The method according to claim 2, wherein, The intelligent agent system also includes a script compliance review intelligent agent; The method further includes: The script compliance review agent is invoked to conduct a compliance review of the short video script. If the review fails, modification suggestions are generated. The proposed modifications are injected into the input context of the creative agent, guiding the creative agent to regenerate the short video script.

5. The method according to any one of claims 2-4, wherein, The output of the director's intelligent agent also includes character images and storyboards; The intelligent agent system also includes a visual compliance review intelligent agent; The method further includes: The visual compliance review agent is invoked to conduct a compliance review of the character image and storyboard. If the review fails, adjustment suggestions are generated. The adjustment suggestions are injected into the input context of the director agent, guiding the director agent to regenerate the character image and storyboard.

6. The method according to claim 2, wherein, The product knowledge also includes high-conversion material features, which are the script structure features of short videos extracted from short video materials with high conversion rates related to the target insurance product. The dialogue rhythm and the timing of introducing the target insurance product in the short video script are generated by the creative agent with reference to the characteristics of the high-conversion materials.

7. The method according to claim 6, wherein, The knowledge base includes a high-conversion material feature library, which includes multiple script structure features extracted from high-conversion short video materials. The method further includes: Obtain feedback data on the short videos after they are deployed; High-conversion short video materials are identified based on the feedback data; Script structure features were extracted from the high-conversion short video material; The high-conversion material feature library is updated based on the extracted script structure features.

8. A short video generation device, comprising: The acquisition module is used to acquire the user's input description of the target short video to be generated; The recall module is used to recall product knowledge related to the target insurance product from the knowledge base based on the requirement description. The product knowledge includes the insured user profile, the accident scenario, and the relationship between the insured and the policyholder. The generation module is used to take the requirement description and the product knowledge as the input context of the intelligent agent system, and guide the intelligent agent system to generate a target short video. The character images in the target short video are generated by the intelligent agent system with reference to the insured user profile, the plot of the target short video is generated by the intelligent agent system with reference to the accident scenario, and the relationships and dialogues between the characters in the target short video are generated by the intelligent agent system with reference to the relationship between the insured and the policyholder.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-7.

10. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-7.

Images