Method and device for realizing self-evolution of video generation agent by computing power through intelligent computing cloud platform

Abstract

The application provides a method and device for realizing self-evolution of a video generation intelligent agent by computing power through an intelligent computing cloud platform, and relates to the technical fields of intelligent computing centers, intelligent computing centers, computing power infrastructure and intelligent computing cloud technology.The method comprises the following steps: S1, obtaining a video generation requirement and a business rule; S2, inputting the requirement into a video generation intelligent agent to output a structured generation action; S3, generating a candidate video according to the structured generation action; S4, obtaining an evaluation result of each candidate video based on the business rule; S5, converting the evaluation result into a reward by using a reward compiler; S6, updating an intelligent agent strategy according to the rewards and the structured generation actions corresponding to all candidate videos; and S7, performing version regression verification on the updated strategy, and releasing the strategy when the verification result meets the release condition. The strategy self-evolution of the video generation intelligent agent is realized, and the stability of the generated video effect is improved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent computing centers, smart computing centers, computing infrastructure, and smart cloud computing technologies, specifically to a method and apparatus for an intelligent computing cloud platform to achieve self-evolution of video-generated intelligent agents through computing power. Background Technology

[0002] With the rapid development of artificial intelligence technology, "intelligent computing centers" and "smart computing centers" have emerged.

[0003] An "intelligent computing center" refers to a facility that provides the necessary computing power, data, and algorithms for artificial intelligence applications (such as the development, training, and inference of deep learning models) by utilizing large-scale heterogeneous computing resources, including general-purpose and intelligent computing power. Intelligent computing centers encompass facilities, hardware, and software, and can provide full-stack capabilities from underlying computing power to top-level application enablement.

[0004] "Intelligent computing center" includes, but is not limited to, "intelligent computing center".

[0005] "Intelligent computing center" or artificial intelligence computing center is a type of computing infrastructure that provides computing power services, data services, and algorithm services required for artificial intelligence applications, based on artificial intelligence theory and adopting artificial intelligence computing architecture.

[0006] "Computing power" is the core of "intelligent computing center" and "smart computing center". It is the ability of computer equipment or computing / data center to process parameters. It is the ability of computer hardware and software to work together to execute a certain computing requirement. It is the computing power to achieve the target result output by processing parameter data. It is a new type of productivity that integrates parameter computing power, network carrying capacity and data storage capacity. It mainly provides services to society through computing power infrastructure.

[0007] In recent years, multimodal video generation technologies such as text-based video, image-based video, audio-driven video, and character animation have developed rapidly. Video generation models are able to generate short videos with a certain degree of realism, motion continuity, and semantic consistency at higher resolutions and frame rates. As the capabilities of generation models improve, the production chain is gradually shifting from "getting a video from a single input prompt" to an intelligent workflow where "users submit creative requirements, the system automatically understands the requirements, plans shots, generates scripts, organizes storyboards, calls multiple models, and outputs a deliverable video."

[0008] However, while the intelligent computing center provides computational support for the video generation agent, the challenge in real-world video generation scenarios is not merely whether the agent can output high-quality videos, but rather whether the system can consistently and stably generate videos that meet user intent and business rules. Simply relying on a single prompt word project or manual parameter tuning is insufficient to consistently cover the quality requirements across different themes, styles, durations, resolutions, and business rules. Currently, the evaluation of generated videos typically remains at the result assessment level, used only for sorting, filtering, reporting, or manual review, and is unlikely to drive continuous improvement in the video generation agent.

[0009] It is evident that since the emergence of intelligent computing centers, video generation agents have lacked a mechanism to use evaluation results to optimize generation strategies. This results in high costs and poor stability of video generation, a problem that remains to be solved in this field. Summary of the Invention

[0010] This invention provides a method and apparatus for enabling the self-evolution of video generation intelligence through computing power on an intelligent computing cloud platform. This addresses the problems in existing technologies where video generation intelligence lacks a mechanism to optimize generation strategies using evaluation results, resulting in high costs and poor performance stability in video generation.

[0011] To solve the above problems, the present invention is implemented as follows: In a first aspect, the present invention provides a method for an intelligent computing cloud platform to achieve self-evolution of a video-generating intelligent agent through computing power, comprising: Step S1: Obtain user video generation requirements and corresponding business rules; Step S2: Input the user's video generation requirements into the video generation agent, schedule the computing power of the intelligent computing cloud platform, and obtain at least one set of structured generation actions output by the video generation agent; Step S3: Based on the at least one set of structured generation actions, schedule the computing power of the intelligent computing cloud platform to generate at least one candidate video; Step S4: Based on the business rules, perform multi-dimensional video evaluation on each candidate video to obtain the evaluation results; Step S5: Using a reward compiler, the evaluation results are converted into reinforcement learning rewards; Step S6: Update the strategy of the video generation agent according to the reinforcement learning reward and structured generation action corresponding to each candidate video; Step S7: Perform version regression verification on the updated strategy of the video generation agent, and release the updated strategy when the verification result meets the release conditions to obtain the self-evolved video generation agent.

[0012] In one embodiment, step S4 includes: Step S4.1: Perform at least two of the following on the candidate video: objective quality evaluation, multimodal semantic evaluation, text alignment evaluation, originality evaluation, rule adaptation evaluation, cost evaluation, and security evaluation, to obtain the evaluation vector; Step S4.2: Determine the cost required to generate the candidate video; Step S4.3: Based on the hard threshold class rules included in the business rules, determine the hard threshold result corresponding to the candidate video. The hard threshold result includes a parameter value indicating whether the hard threshold class rule is satisfied, and a value indicating the degree of violation of the hard threshold class rule. Step S4.4: Perform evidence consistency verification on the candidate videos to obtain evidence consistency scores; Step S4.5: Based on the user's preference data, determine the preference gain corresponding to the candidate video; Step S4.6: Obtain the evaluation result from the evaluation vector, the cost, the hard threshold result, the evidence consistency score, and the preference gain.

[0013] In one embodiment, step S5 includes: Step S5.1: Based on the weight of each component in the evaluation vector in the business rules, perform a weighted summation of all components to obtain the basic reward; Step S5.2: Based on the evidence consistency score, determine the abnormal speculation risk value, which is used to describe the penalty index in the reward construction of the video generation agent reinforcement learning; Step S5.3: Combine the basic reward with the abnormal speculative risk value, the cost, the degree value, and the preference gain to calculate the reinforcement learning reward.

[0014] In one embodiment, after step S5.3, step S5 further includes: Step S5.4: Based on the parameter values ​​in the hard threshold result, the reinforcement learning reward, and the preset failure penalty, determine the gating reward for the candidate video, and use the gating reward to update the strategy of the video generation agent.

[0015] In one embodiment, step S6 includes: Step S6.1: Determine the relative advantage of each candidate video group based on the gating rewards corresponding to all candidate videos. Step S6.2: Update the strategy of the video generation agent based on the relative advantages of the groups corresponding to the candidate videos and the structured generation actions.

[0016] In one embodiment, step S6 includes: Step S6.1': For any two candidate videos, construct preference samples based on the reinforcement learning rewards and structured generation actions corresponding to the two candidate videos respectively; Step S6.2': Update the strategy of the video generation agent based on the preference sample.

[0017] In one embodiment, after step S5, the method further includes: Step S8: Based on the pre-built action dependency graph, determine the attribution weight corresponding to each action in the structured generated action, wherein the action dependency graph contains the mapping relationship between multiple actions and video evaluation indicators; Step S9: Determine the local reward corresponding to the action based on the reinforcement learning reward of the candidate video, the attribution weight, and the score corresponding to the action in the evaluation result.

[0018] In one embodiment, prior to step S5, the method further includes: Step S10: Parse, clean and structure the business rules to obtain an item set. The item set contains multiple items, and each item contains at least one of the following: tag, description, tier, weight, hard threshold attribute, reward mapping method and applicable scenario. Step S11: Based on the set of entries, form the reward configuration of the reward compiler.

[0019] In one embodiment, the structured generation action includes model selection, and step S3 includes: Step S3.1: Based on the model selection, determine at least one model for generating the video; Step S3.2: Invoke the at least one model and generate candidate videos based on the structured generation action.

[0020] In one embodiment, prior to step S2, the method further includes: Step S12: Based on the current task budget of the self-evolution of the video generation agent, determine the computing power orchestration strategy of the intelligent computing cloud platform. The computing power orchestration strategy controls at least one of the following: the number of candidate videos generated, generation parameters, evaluation dimensions, and training frequency of the video generation agent.

[0021] Secondly, the present invention also provides a device for an intelligent computing cloud platform to achieve self-evolution of video-generated intelligent agents through computing power, comprising: The acquisition module is used to acquire user video generation requirements and corresponding business rules; The intelligent agent module is used to input the user's video generation requirements into the video generation intelligent agent, schedule the computing power of the intelligent computing cloud platform, and obtain at least one set of structured generation actions output by the video generation intelligent agent. The video generation module is used to schedule the computing power of the intelligent computing cloud platform to generate at least one candidate video based on the at least one set of structured generation actions. The evaluation module is used to perform multi-dimensional video evaluation on each candidate video based on the business rules and obtain the evaluation results; The reward compilation module is used to transform the evaluation results into reinforcement learning rewards using a reward compiler. The update module is used to update the strategy of the video generation agent based on the reinforcement learning rewards and structured generation actions corresponding to all candidate videos. The verification module is used to perform version regression verification on the updated strategy of the video generation agent, and release the updated strategy when the verification result meets the release conditions, so as to obtain the self-evolved video generation agent.

[0022] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, it implements the steps in the method for achieving self-evolution of video-generating intelligent agents through computing power by the intelligent computing cloud platform described in the first aspect above.

[0023] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the method for achieving self-evolution of a video-generating intelligent agent through computing power by an intelligent computing cloud platform as described in the first aspect above.

[0024] Fifthly, the present invention also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps in the method for achieving self-evolution of a video-generating intelligent agent through computing power by an intelligent computing cloud platform as described in the first aspect above.

[0025] In this invention, step S1 is to obtain user video generation requirements and corresponding business rules; step S2 is to input user video generation requirements into a video generation agent, schedule the computing power of the intelligent computing cloud platform, and obtain at least one set of structured generation actions output by the video generation agent; step S3 is to schedule the computing power of the intelligent computing cloud platform based on at least one set of structured generation actions to generate at least one candidate video; step S4 is to perform multi-dimensional video evaluation on each candidate video based on business rules and obtain evaluation results; step S5 is to use a reward compiler to convert the evaluation results into reinforcement learning rewards; step S6 is to update the strategy of the video generation agent based on the reinforcement learning rewards and structured generation actions corresponding to all candidate videos; step S7 is to perform version regression verification on the updated strategy of the video generation agent, and release the updated strategy when the verification results meet the release conditions, thus obtaining a self-evolved video generation agent. In this way, by obtaining at least one set of structured generation actions from the video generation agent in response to user needs, generating multiple candidate videos using these structured generation actions, evaluating each candidate video, compiling the evaluation results into reinforcement learning rewards, updating the video generation agent's strategy based on the reinforcement learning rewards and structured generation actions, and performing version regression verification on the updated strategy, releasing it after successful verification, the video generation agent's strategy can achieve self-evolution, greatly improving the stability of videos generated by computing power in the intelligent computing cloud platform, improving video quality, better meeting user needs and business rules, and reducing the cost of video generation. Attached Figure Description

[0026] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the present invention will be briefly introduced below. Obviously, the accompanying 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.

[0027] Figure 1 This is a flowchart of a method for achieving self-evolution of video-generating intelligent agents through computing power using an intelligent computing cloud platform, as provided by this invention. Figure 2 This is a structural diagram of a device provided by the present invention that enables the self-evolution of video-generating intelligent agents through computing power in an intelligent computing cloud platform. Figure 3 This is a structural diagram of an electronic device provided by the present invention. Detailed Implementation

[0028] The technical solutions of this invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] The “computing power” mentioned in this invention refers to: the ability of computer equipment or computing / data center to process information; the ability of computer hardware and software to work together to perform a certain computing requirement; the computing power to achieve the target result output by processing information data; and a new type of productivity that integrates information computing power, network carrying capacity, and data storage capacity, mainly providing services to society through computing power infrastructure.

[0030] The "computational power" (CP) described in this invention refers to the ability of a data center server to process data and output results. It is a comprehensive indicator of a data center's computing power, encompassing general computing power, supercomputing power, and intelligent computing power. The commonly used unit of measurement is floating-point operations per second (FLOPS, 1 EFLOPS = 10^18 FLOPS), with higher values ​​indicating stronger overall computing power. It is estimated that 1 EFLOPS is approximately the computing power output of 5 Tianhe-2A supercomputers, 500,000 mainstream server CPUs, or 2 million mainstream laptops. The calculation formula is: CP = CP 通用 +CP 智能 +CP 超级 .

[0031] The "Network Power" (NP) mentioned in this invention refers to the performance of data transmission capability of computing facilities, which includes comprehensive capabilities such as network architecture, network bandwidth, transmission latency, intelligent management and scheduling, and involves network transmission within and between data centers. It is a comprehensive indicator for measuring network transmission scheduling capability.

[0032] The "Storage Power" (SP) described in this invention refers to the comprehensive capabilities of a data center in four aspects: data storage capacity, performance, security and reliability, and green and low-carbon operation. It is a comprehensive indicator for measuring the data storage capacity of a data center, including external storage devices such as storage arrays and internal storage devices within servers. The commonly used unit of measurement for storage capacity is exabytes (EB, 1EB = 2^60 bytes), while the commonly used unit of measurement for performance is the number of read / write operations per second (IOPS / TB). Disaster recovery ratio is an important indicator of security and reliability.

[0033] The "computing infrastructure" mentioned in this invention refers to a new type of information infrastructure that integrates information computing power, network carrying capacity, and data storage capacity, enabling centralized computing, storage, transmission, and application of information.

[0034] The "new information infrastructure" mentioned in this invention refers to network infrastructure such as 5G networks, fiber optic broadband networks, backbone networks, international communication networks, and satellite internet; computing infrastructure such as data centers, general computing centers, intelligent computing centers, and supercomputing centers; and new technology facilities such as artificial intelligence, blockchain, and quantum computing.

[0035] The “computing power” mentioned in this invention includes: general computing power, intelligent computing power, and supercomputing power.

[0036] The "general computing power" mentioned in this invention refers to the computing power provided by servers based on CPU (Central Processing Unit) chips, which is used to support basic general computing such as cloud computing and edge computing.

[0037] The "intelligent computing power" mentioned in this invention refers to: a computing platform deployed on a large scale based on dedicated chips such as GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), and ASIC (Application Specific Integrated Circuit) for various artificial intelligence innovative applications, such as natural language processing and machine vision.

[0038] The “supercomputing power” mentioned in this invention refers to the computing power provided by high-performance computing clusters such as supercomputers. It utilizes the centralized computing resources of multiple computer systems working in parallel and uses a dedicated operating system to handle extremely complex or data-intensive problems. It is mainly used for computing in cutting-edge scientific fields, such as planetary simulation, drug molecule design, and gene analysis.

[0039] The "intelligent computing center" described in this invention refers to a facility that, through the use of large-scale heterogeneous computing resources, including general-purpose computing power (CPU) and intelligent computing power (GPU, FPGA, ASIC, etc.), primarily provides the necessary computing power, data, and algorithms for artificial intelligence applications (such as the development, training, and inference of deep learning models). The intelligent computing center encompasses facilities, hardware, and software, and can provide full-stack capabilities from underlying computing power to top-level application enablement.

[0040] The "intelligent computing cloud platform" mentioned in this invention, abbreviated as "intelligent computing cloud", refers to a cloud computing platform that integrates hardware and software resources based on an intelligent computing center.

[0041] The "intelligent computing center" mentioned in this invention includes, but is not limited to, "smart computing center".

[0042] The "intelligent computing center" mentioned in this invention, also known as an artificial intelligence computing center, is a type of computing infrastructure that provides computing power services, data services, and algorithm services required for artificial intelligence applications, based on artificial intelligence theory and adopting an artificial intelligence computing architecture.

[0043] The "computing center" mentioned in this invention refers to a facility that is mainly composed of infrastructure such as wind, thermal, hydro, and electricity, and IT hardware and software equipment, and has computing power, carrying capacity, and storage capacity, including general data centers, intelligent computing centers, supercomputing centers, etc.

[0044] The "supercomputing center" mentioned in this invention refers to a supercomputing data center, which is a data center based on supercomputers or large-scale computing clusters. It can provide large-scale computing, storage and network services and is widely used in aerospace, defense, oil exploration, climate modeling and genome sequencing and other application scenarios.

[0045] The “computing resources” mentioned in this invention refer to the technologies and facilities required for the development of the digital society that have the ability to compute, transmit, store and apply information, including but not limited to computing resources such as CPUs and GPUs, network resources such as switches and routers, storage resources such as storage arrays and distributed storage, security resources such as firewalls and intrusion detection systems, and supporting and guaranteeing resources such as wind, fire, water and electricity.

[0046] The "model" mentioned in this invention includes, but is not limited to, "large language model" and "multimodal large model".

[0047] The "large language model" mentioned in this invention refers to a large-scale language model (LLM), which is a language model with a large number of parameters. It is designed to understand and generate human language, and is trained with a large amount of text data. It can perform a wide range of tasks, including text summarization, translation, and sentiment analysis.

[0048] The “Multimodal Large Models” mentioned in this invention refer to models that combine multimodal information such as text, images, videos, and audio for training, including but not limited to multimodal large language models.

[0049] The "agent" described in this invention refers to an agent capable of perceiving the environment and taking actions to achieve specific goals. It can be software, hardware, or a system, possessing autonomy, adaptability, and interactivity. The agent perceives changes in the environment (e.g., through sensors or data input), makes judgments and decisions based on its learned knowledge and algorithms, and then executes actions to influence the environment or achieve predetermined goals.

[0050] Please see Figure 1 , Figure 1 This is a flowchart of a method for achieving self-evolution of video-generating intelligent agents through computing power using an intelligent computing cloud platform, as provided by this invention. Figure 1 As shown, the method includes: Step S1: Obtain user video generation requirements and corresponding business rules.

[0051] The video generation requirements mentioned above may include, but are not limited to, at least one of the following: theme, characters, scene, style, duration, shot, resolution, and reference materials.

[0052] The aforementioned business rules may include, but are not limited to, external rules, competition regulations, brand requirements, platform review standards, or internal review experience.

[0053] The beneficial technical effects achieved by step S1 are as follows: By synchronously acquiring business rules and user generation requirements, the intelligent agent can refer to these constraints in real time during the generation process, thereby ensuring the synchronous guarantee of generation quality and compliance. This provides a data foundation for the self-evolution results of the video generation intelligent agent in the intelligent computing cloud platform to better meet user and business needs.

[0054] Step S2: Input the user's video generation requirements into the video generation agent, schedule the computing power of the intelligent computing cloud platform, and obtain at least one set of structured generation actions output by the video generation agent.

[0055] The aforementioned video generation agent is not limited to a single text-to-video model, but rather includes a combination of at least one or more modules such as user needs analysis, plot / script generation, storyboard planning, cue word expansion, reference image selection, basic video generation model invocation, sampling parameter selection, post-processing, quality evaluation, reward attribution, strategy update, and version release, which can provide reliable action guidance for the model to generate videos.

[0056] The aforementioned structured generation actions may include at least one of the following: plot planning, storyboard planning, cue words, model selection, generation parameters, or post-processing parameters.

[0057] In this invention, the video-generating agent does not output a prompt word, but rather outputs a set of structured actions as shown in formula (1): (1) In formula (1), For the video generation requirement, the current policy of the video generation agent is denoted as . , Indicates plot or script planning. Indicates the storyboard planning, This indicates a positive prompt or video generation description. Indicates a negative indicator word. This indicates the choice of a base video generation model or a combination of models. This refers to generation parameters such as sampling steps, resolution, frame rate, random seed, guidance coefficient, and motion intensity. This indicates post-processing parameters such as frame interpolation, super-resolution, artifact removal, cropping, dubbing, or subtitles.

[0058] In one embodiment, the video generation agent can appropriately adjust one or more of the structured generation actions shown in formula (1) to obtain multiple sets of structured generation actions for the same user's video generation needs.

[0059] In one embodiment, the base video generation model can be a diffusion transformer model, a stream matching model, a hybrid expert video model, a text-to-video model, an image-to-video model, an audio-driven video model, or other models capable of generating video content. Alternatively, multiple base video generation models can be combined to jointly generate candidate videos.

[0060] In some possible implementations, the structured generation action may also include a retry strategy.

[0061] The beneficial technical effects achieved by step S2 are as follows: By generating multiple sets of structured generation actions for the same user video generation needs, the parallel computing power advantage of the intelligent computing cloud platform can be fully utilized to increase the diversity of actions on which video generation is based, provide high-quality comparative learning signals for reinforcement learning, and significantly improve the output value of a single computing power investment.

[0062] Step S3: Based on at least one set of structured generation actions, schedule the computing power of the intelligent computing cloud platform to generate at least one candidate video.

[0063] In one embodiment, a video generation model can be invoked to process structured generation actions, thereby generating candidate videos that match the plot, scene composition, resolution, subtitles, etc., in the structured generation actions. A set of structured generation actions is obtained based on step S2 above. Then, candidate videos can be generated using the following formula (2).

[0064] (2) In formula (2), This indicates the basic video generation model or model service that can be accessed. This indicates random noise, seed, or sampling perturbation.

[0065] In one embodiment, since the cost of generating a video once is high, the score of a single candidate video is easily affected by sampling randomness. Therefore, in order to reduce scoring noise, the present invention can generate multiple candidate videos (such as K) under the same user demand. The candidate set composed of these multiple candidate videos can be represented by the following formula (3): (3) in, =1, 2, ..., K.

[0066] It should be noted that for the above formula (3), different structured generation actions can choose different models or combinations of models to generate candidate videos.

[0067] The beneficial technical effects achieved by step S3 are as follows: by generating multiple candidate videos based on multiple sets of structured generation actions, the quality and diversity of video generation are improved, which provides sufficient basis for the self-evolution of video generation agents on the intelligent computing cloud platform and is conducive to improving the efficiency of self-evolution.

[0068] Step S4: Based on business rules, perform multi-dimensional video evaluation on each candidate video to obtain the evaluation results.

[0069] The aforementioned multidimensional video evaluation may include, but is not limited to, at least two of the following: objective quality evaluation, multimodal semantic evaluation, text alignment evaluation, originality evaluation, rule adaptation evaluation, cost evaluation, and security evaluation.

[0070] In one embodiment, objective quality assessment focuses on the technical specifications of the video as a media file itself, and the assessment content may include resolution, frame rate, bit rate, image clarity, motion smoothness, etc.

[0071] Multimodal semantic evaluation focuses on the semantic consistency and logical rationality of videos across multiple modalities, including visual, text, and audio. Evaluation content may include whether the visual content is semantically aligned with the text prompts, whether the audio and visuals semantically match (e.g., if a piano is played on screen, is the audio indeed the sound of a piano), whether character actions conform to physical principles, and whether scene transitions are logically coherent. Text alignment evaluation assesses the degree of semantic matching between the generated video and the input text prompts.

[0072] The originality assessment evaluates the similarity between the generated video and known works in the assessment dataset, publicly available videos on the Internet, and other candidate videos in the same batch. This can include similarity in composition, camera language, and musical melody.

[0073] Cost evaluation focuses on the computing resources and economic costs consumed in generating the candidate video. Evaluation content may include, but is not limited to, generation costs (such as GPU time and video memory usage), evaluation costs for assessing model inference overhead, and storage costs for video files and intermediate products.

[0074] Rule adaptation assessment focuses on the degree to which the generated video complies with specific business rules, while security assessment evaluates whether candidate videos violate the hard threshold conditions in the business rules, such as containing illegal content or exceeding the prescribed duration.

[0075] The beneficial technical effect achieved by step S4 is that by performing multi-dimensional evaluation of the video, more refined evaluation results are obtained, which can provide accurate reward signals for reinforcement learning.

[0076] Step S5: Use a reward compiler to convert the evaluation results into reinforcement learning rewards.

[0077] The aforementioned reward compiler is an automated rule-to-reward transformation engine that can translate business rules into reward configurations that reinforcement learning algorithms can directly load and execute.

[0078] In one embodiment, the reward compiler can transform the evaluation results, including the evaluation score, hard threshold result, failure reason, cost metric, and evidence consistency, into reward vectors and scalar rewards that can be used for reinforcement learning.

[0079] It should be noted that in this invention, the reward can also be generated by a rule function, multimodal large model scoring, manual review, preference model, video quality model, security model, or a combination thereof. In the intelligent computing cloud platform, either an explicit weighted formula or a learned reward model can be used; either a single scalar reward can be output, or a multi-objective reward vector can be output and selected using a Pareto policy.

[0080] The beneficial technical effects achieved in step S5 are as follows: by using a reward compiler, the multidimensional evaluation results are transformed into reinforcement learning reward signals, thereby achieving unified quantization of heterogeneous evaluation signals and ensuring the stability of the self-evolution of video generation agents in the intelligent computing cloud platform.

[0081] Step S6: Update the policy of the video generation agent based on the reinforcement learning rewards and structured generation actions corresponding to all candidate videos.

[0082] The strategy update for the aforementioned video generation agent can optimize only the prompt word strategy, or it can optimize the scene planning, model selection, sampling parameters, negative prompt words, post-processing parameters, LoRA adapter, or control network. For closed-source models, black-box strategy optimization is the primary approach; for open-source or proprietary models, a lightweight adapter or some model parameters can be further trained.

[0083] In this invention, the policy of the video generation agent can be updated using group relative policy optimization, proximal policy optimization, preference optimization, reward-weighted regression or other reinforcement learning methods. As long as the video generation agent is updated using video evaluation results or rewards compiled from the evaluation results, it falls within the scope of this invention.

[0084] In one embodiment, the structured generation actions among all candidate videos can be determined based on the reinforcement learning reward size corresponding to each candidate video. This allows the strategy of the video generation agent to be updated, making the agent more inclined to generate these structured generation actions, which serve as the basis for the model to generate videos.

[0085] The beneficial technical effect achieved by step S6 is that it enables a closed loop of reinforcement learning from behavior to result to improvement, so that the agent no longer statically executes preset rules, but can extract "what kind of action combination can bring higher rewards in what kind of scenario" from each generated experience, and internalize this experience into the parameter update of the policy network, so as to achieve continuous accumulation of capabilities.

[0086] Step S7: Perform version regression verification on the updated strategy of the video generation agent, and release the updated strategy when the verification results meet the release conditions to obtain the self-evolved video generation agent.

[0087] In this invention, the business rule entries Rubric, reward configuration, agent strategy, model adapter, regression sample set, and release status can be incorporated into unified version governance. After each update of the video generation agent strategy, the intelligent computing cloud platform can verify it on a fixed regression set and obtain the version benefit function as shown in the following formula (4): (4) In formula (4), This indicates an effective improvement in quality. This indicates a change in the dispute rate or failure rate. Indicates cost changes, This indicates a change in security / hard threshold violations. The updated strategy will only be released in a canary release or official release when the value of the version's benefit function meets preset conditions.

[0088] In some possible embodiments, user requirements, action trajectories, candidate video summaries, evaluation results, rewards, manual review conclusions, and version information can also be saved. The Rubric version, reward version, strategy version, model adapter version, regression set, and canary release status can be managed. If quality degradation, abnormal costs, or increased disputes occur during online operation, the intelligent computing cloud platform can roll back to any stable historical version.

[0089] The beneficial technical effects achieved by step S7 are: through version governance and regression verification, rule changes, reward changes and strategy changes can be audited, rolled back and reviewed, thereby improving the reliability of the intelligent computing cloud platform in realizing the self-evolution of intelligent agents.

[0090] In this invention, step S1 is to obtain user video generation requirements and corresponding business rules; step S2 is to input user video generation requirements into a video generation agent, schedule the computing power of the intelligent computing cloud platform, and obtain at least one set of structured generation actions output by the video generation agent; step S3 is to schedule the computing power of the intelligent computing cloud platform based on at least one set of structured generation actions to generate at least one candidate video; step S4 is to perform multi-dimensional video evaluation on each candidate video based on business rules and obtain evaluation results; step S5 is to use a reward compiler to convert the evaluation results into reinforcement learning rewards; step S6 is to update the strategy of the video generation agent based on the reinforcement learning rewards and structured generation actions corresponding to all candidate videos; step S7 is to perform version regression verification on the updated strategy of the video generation agent, and release the updated strategy when the verification results meet the release conditions, thus obtaining a self-evolved video generation agent. In this way, by obtaining at least one set of structured generation actions from the video generation agent in response to user needs, generating multiple candidate videos using these structured generation actions, evaluating each candidate video, compiling the evaluation results into reinforcement learning rewards, updating the video generation agent's strategy based on the reinforcement learning rewards and structured generation actions, and performing version regression verification on the updated strategy, releasing it after successful verification, the video generation agent's strategy can achieve self-evolution, greatly improving the stability of videos generated by computing power in the intelligent computing cloud platform, improving video quality, better meeting user needs and business rules, and reducing the cost of video generation.

[0091] In other words, the intelligent computing cloud platform provided by this invention enables the self-evolution of video generation agents through computing power. This allows video evaluation to move beyond being a static scoring tool at the end of the generation chain and become the basis for reward construction, reward attribution, and policy updates in the reinforcement learning loop. As a result, the video generation agent can continuously improve generation quality, text alignment, plot expression, motion naturalness, business pass rate, and computing power utilization efficiency during continuous task execution.

[0092] For example, in a video generation business scenario, the user input might be "Generate a 5-second short video of a robot assisting a doctor in providing first aid in a futuristic city." The user input can be analyzed first, extracting the subject, scene, actions, style, duration, and safety requirements. The video generation agent then generates K=4 sets of structured action plans for the same requirement, including different plot focuses, shot order, cue details, negative cue words, and sampling parameters. Afterward, the basic video generation module can be called to output candidate videos for each of the four sets of structured action plans.

[0093] Then, a unified evaluation can be performed on the four candidate videos. The first candidate video has good image quality, but the relationship between the robot and the doctor's movements is unclear; the second candidate video has natural movements, but lacks elements of the emergency rescue scene; the third candidate video has good text alignment, character relationships, camera stability, and creative expression; the fourth candidate video has screen flickering and abnormal hand movements. Afterwards, the reward compiler calculates the rewards for each of the four candidate videos based on objective metrics, semantic metrics, and hard threshold penalties, and calculates the relative advantage of each group.

[0094] Then, based on the relative advantages of the groups, when updating the video generation agent, the script planning, storyboard planning, cue word expansion, and sampling parameters corresponding to the third candidate video can be used as positive reinforcement objects, while the high flicker parameter combinations and unstable action descriptions in the fourth candidate video can be used as negative samples. After accumulating data from multiple batches of tasks, the video generation agent gradually learns that: in medical emergency videos, the main character, key actions, camera distance, and environmental information should be clearly defined; for complex interactive scenes, the amplitude of movement and camera switching should be controlled; and for tasks prone to limb abnormalities, negative constraints should be added. As a result, the text alignment, motion naturalness, and business pass rate of subsequent similar videos are improved.

[0095] In one embodiment, step S4 includes: Step S4.1: Perform at least two of the following on the candidate video: objective quality assessment, multimodal semantic assessment, text alignment assessment, originality assessment, rule adaptation assessment, cost assessment, and security assessment, to obtain the assessment vector.

[0096] The aforementioned objective quality assessment refers to evaluating video image quality, temporal stability, and motion quality, which may include, but is not limited to, Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Video Multimethod Assessment Fusion (VMAF), sharpness, optical flow smoothness, jitter rate, artifacts, noise, audio-visual synchronization, lip-sync, subject integrity, and shot transition stability.

[0097] The aforementioned multimodal semantic evaluation refers to the evaluation of plot integrity and expression, which may include, but is not limited to, text alignment, plot integrity, conflict progression, character consistency, and style consistency.

[0098] In this invention, for candidate videos The resulting evaluation vector can be represented by the following formula (5): (5) In formula (5), Indicates objective image quality score. Indicates time-stable components, Indicates the motion mass score. Indicates audio-visual synchronization. This indicates text / scene alignment. This indicates the completeness of the plot and the level of expression. Indicates the creativity and originality score. This indicates the business rule adaptation score. Indicates safety and compliance score. This indicates cost delay or resource efficiency.

[0099] In some possible implementations, objective image quality, temporal stability, and motion quality can inherit normalized indicators from existing video evaluation systems, such as the following formula (6).

[0100]

[0101] (6) Semantic evaluation metrics are output by a Vision-Language Model (VLM) or a multimodal evaluation model based on structured Rubric, retaining the reasoning, keyframe evidence, and failure cause for each score. The intelligent computing cloud platform does not directly use free text as a reward; instead, it first normalizes it into structured numerical values ​​and discrete labels.

[0102] Step S4.2: Determine the cost required to generate candidate videos.

[0103] In one embodiment, the value corresponding to the cost assessment in step S4.1 in the assessment vector can be determined as the cost required for candidate video generation; or existing cost assessment models or tools can be used to calculate the computing power consumed by video generation and determine the cost required for candidate video generation.

[0104] Step S4.3: Determine the hard threshold result corresponding to the candidate video based on the hard threshold rules included in the business rules.

[0105] The aforementioned hard threshold results can include parameter values ​​indicating whether a hard threshold rule is met, and a degree value indicating the extent of violation. Parameter values ​​can be 0 or 1. A parameter value of 0 indicates that the candidate video does not comply with the hard threshold rules included in the business rules, such as containing illegal content or exceeding a specified video duration threshold; conversely, a parameter value of 1 indicates that the candidate video fully complies with the hard threshold rules. The degree value can be understood as, for example, the length of time exceeding the duration limit.

[0106] In one embodiment, due to differences in event regulations, brand requirements, or platform review standards, the hard threshold rules included in different business rules may also be different. For example, the hard threshold rules included in the business rules corresponding to a certain event may be that the video resolution must reach a specific resolution, the video duration cannot exceed a certain duration, etc.

[0107] Step S4.4: Perform evidence consistency verification on the candidate videos to obtain evidence consistency scores.

[0108] In one embodiment, to prevent agents from speculatively increasing their scores, this invention establishes a reward-based anti-cheating mechanism. The intelligent computing cloud platform can perform evidence consistency checks on candidate videos, including consistency between keyframe evidence and VLM descriptions, consistency between shot segments and storyboard scenes, consistency between audio waveforms and lip movements / actions, consistency between text entities and image entities, and consistency between objective indicators and semantic evaluations, thereby obtaining an evidence consistency score.

[0109] In this invention, evidence consistency verification of candidate videos can suppress phenomena such as excessive pandering to scorers, low-diversity repetition, and rule boundary avoidance.

[0110] Step S4.5: Determine the preference gain corresponding to the candidate video based on the user's preference data.

[0111] In one embodiment, preference data obtained from statistics on users' historical video browsing, such as preferred video duration, preferred plot, and preferred scene composition, can be matched with generated candidate videos. The higher the degree of matching, the greater the preference gain corresponding to the candidate video.

[0112] It should be noted that in this invention, diversity gain can also be calculated based on the diversity of valid content in the candidate videos.

[0113] Step S4.6: Obtain the evaluation results from the evaluation vector, cost, hard threshold results, evidence consistency score, and preference gain.

[0114] The beneficial technical effects achieved by steps S4.1 to S4.6 are: the quality of the generated video can be evaluated from multiple dimensions, ensuring the comprehensiveness of the evaluation results, providing a reliable reinforcement learning basis for the intelligent computing cloud platform to realize the self-evolution of intelligent agents, and helping to improve the efficiency of computing power utilization.

[0115] In one embodiment, the reward compiler can compile the evaluation vector, hard threshold, cost term, and evidence consistency into a reinforcement learning reward. After obtaining the evaluation results based on steps S4.1 to S4.6 above, step S5 includes: Step S5.1: Based on the weight of each component in the evaluation vector in the business rules, perform a weighted sum of all components to obtain the basic reward.

[0116] In this invention, the basic reward It can be calculated using the following formula (7): (7) In formula (7), This is an indicator mapping function used to transform evaluation results with different dimensions and forms into a unified reward. For example, a step penalty is used for hard threshold indicators, a smooth mapping is used for image quality indicators, a strong penalty is used for security indicators, and a negative reward is used for cost indicators.

[0117] Step S5.2: Determine the risk value of abnormal speculation based on the evidence consistency score.

[0118] The aforementioned anomalous speculative risk value is used to describe the penalty metric in the reward construction of reinforcement learning for video generation agents.

[0119] In one embodiment, based on the evidence consistency score The abnormal speculative risk value can be calculated using the following formula (8). .

[0120] (8) In formula (8), A hard threshold can be set as needed, based on the evidence consistency score. If the value falls below this hard threshold, a linear penalty will be triggered; conversely, if the hard threshold is reached, The item is reset to zero, ensuring that the generated video must first meet the most basic logical and factual requirements before it is eligible to participate in subsequent competitions. This indicates a low risk of duplication due to low diversity. This indicates that an excessive pursuit of a certain objective indicator has led to quality abnormalities. This indicates the boundaries of the rules to avoid risks.

[0121] Step S5.3: Combine the basic reward with the abnormal speculative risk value, cost, degree value and preference gain to obtain the reinforcement learning reward.

[0122] In one embodiment, the complete reinforcement learning reward is calculated by fusing the base reward with the anomalous speculative risk value, cost, severity value, and preference gain. It can be represented by the following formula (9): (9) In formula (9), This represents computational cost and latency cost. Indicates the degree to which hard rules or safety constraints are violated. This indicates the risk value of abnormal speculation. This indicates effective diversity or user preference gain.

[0123] The beneficial technical effects achieved by steps S5.1 to S5.3 are: to avoid intelligent agents pursuing only a single quality score while ignoring issues of cost, security and diversity, to improve the quality and reliability of intelligent agents' self-evolution in intelligent computing cloud platforms, and to improve computing power utilization efficiency.

[0124] It should be noted that, in order to prevent severely substandard videos from receiving positive rewards due to high individual semantic metrics, a hard threshold can be used to further determine the gating reward corresponding to the candidate video. The result of determining the gating reward is used as a new reinforcement learning reward to update the strategy of the video generation agent.

[0125] In one embodiment, after step S5.3, step S5 may further include: Step S5.4: Based on the parameter values ​​in the hard threshold results, the reinforcement learning reward, and the preset failure penalty, determine the gating reward for the candidate video and use the gating reward to update the policy of the video generation agent.

[0126] In this invention, a hard threshold gating system can be applied to address issues such as format, duration, resolution, aspect ratio, security violations, severe audio-visual asynchrony, severe flickering, and obvious generation failures. The parameter values ​​in the hard threshold result... Then gated rewards The calculation method can be shown in the following formula (10): (10) In formula (10), A significant penalty for failure.

[0127] The beneficial technical effect achieved by step S5.4 is that it prevents severely substandard videos from receiving positive rewards due to high individual semantic metrics, thus avoiding contamination of the self-evolution direction by erroneous samples.

[0128] In this invention, a unified evaluation of the candidate set can be performed, and relative advantages can be calculated within the group. Unlike comparing only the absolute scores of individual videos, group-based relative comparison can explicitly amplify the differences between candidates under the same requirements, rules, and computing power budget, making it more suitable for reinforcement learning updates.

[0129] In one embodiment, after obtaining the gating reward corresponding to the candidate video in step S5.4, the group relative advantage of the candidate video can be calculated using the gating reward to update the policy of the video generation agent. Then, step S6 includes: Step S6.1: Determine the relative advantage of each candidate video group based on the gating rewards corresponding to all candidate videos.

[0130] In this invention, the rewards for multiple candidate videos can be standardized under the same user demand to obtain the group relative advantage corresponding to each candidate video, as shown in the following formula (11): (11) Step S6.2: Update the strategy of the video generation agent based on the relative advantages of the groups corresponding to the candidate videos and the structured generation actions.

[0131] In this invention, the strategy of the video generation agent is updated by using the relative advantage of the groups corresponding to candidate videos and the structured generation actions. This allows the intelligent computing cloud platform to focus on "which agent's action is better in the same task" when achieving self-evolution, rather than over-relying on absolute scores across tasks. For video tasks with different themes, styles, and durations, the absolute distribution may differ, while the relative advantage within groups is more stable.

[0132] In one embodiment, the intelligent computing cloud platform can adopt a group-relative strategy with pruning to optimize the objective and update the strategy of the video generation agent, as shown in the following formula (12): (12) In formula (12), , Indicates the reference strategy, This item is used to prevent the strategy from drifting too quickly. Indicates the cropping range. This indicates the constraint strength. This objective can be applied to cue word generation strategies, storyboard planning strategies, model selection strategies, and parameter generation strategies.

[0133] In this invention, for closed-source video generation services that are inconvenient to directly backpropagate, the agent action strategy can be updated through a black-box strategy optimization method.

[0134] For example, in some deployment environments, the basic video generation model is provided as an Application Programming Interface (API) service, making it impossible to directly modify the model parameters. In this case, the present invention can still achieve self-evolution. The intelligent computing cloud platform limits the optimization targets to cue word expansion strategies, storyboard organization strategies, model selection strategies, sampling parameter strategies, and post-processing strategies. After each generation, the intelligent computing cloud platform updates the agent strategy based on evaluation rewards without affecting the basic model parameters. For instance, if the intelligent computing cloud platform discovers that a certain type of motion shot is prone to subject breakage in a certain basic model, the agent strategy will gradually reduce the selection probability of that model in complex motion tasks, or automatically add shot constraints, negative cueing for subject integrity, and more robust sampling parameters. Thus, even if the basic model is not trainable, the intelligent computing cloud platform can still achieve self-evolution through agent action optimization.

[0135] The beneficial technical effects achieved by steps S6.1 to S6.2 are: significantly reducing computing power consumption, enabling more candidate videos to be processed in parallel under the same computing power conditions, significantly improving training throughput, and improving the efficiency of agent self-evolution.

[0136] In one embodiment, in addition to utilizing group relative advantage to update the video generation agent strategy, the present invention can also construct preference pairs and update the video generation agent strategy offline by comparing candidate videos pairwise. Step S6 includes: Step S6.1': For any two candidate videos, construct preference samples based on the reinforcement learning rewards and structured generation actions corresponding to the two candidate videos respectively.

[0137] In some possible embodiments, for any two candidate videos under the same requirement , The corresponding reinforcement learning rewards are denoted as follows: , ,like Greater than That is, satisfy Then you can prefer candidate videos. Corresponding structured generation actions Forming preference samples ,in For optimal action, This is the least desirable action.

[0138] Step S6.2': Update the strategy of the video generation agent based on the preference samples.

[0139] In this invention, after constructing the preference sample, the strategy of the video generation agent can be updated using the following formula (13).

[0140] (13) The beneficial technical effects achieved by steps S6.1' to S6.2' are: to make full use of historical generation and evaluation data and continuously improve the agent's strategy without repeatedly consuming video generation computing power.

[0141] In one embodiment, after step S5, the method further includes: Step S8: Based on the pre-built action dependency graph, determine the attribution weight corresponding to each action in the structured generated actions.

[0142] The action dependency graph above can contain multiple mapping relationships between actions and video evaluation metrics. If there is a mapping relationship between a certain action and a certain video evaluation metric, it means that the magnitude of that video evaluation metric in the video score is affected by that action.

[0143] In one embodiment, the final video score is influenced by the actions of multiple agents. Directly allocating the total reward to all actions may introduce learning noise. Therefore, in this invention, an action dependency graph (labeled as...) can be constructed. This records the impact path of each action on the video result. For example, narrative integrity is mainly related to the script and storyboard, text alignment is mainly related to cue words and model selection, motion naturalness is mainly related to storyboard actions, base model and sampling parameters, and cost is mainly related to model selection, resolution and number of candidates.

[0144] In this invention, the first step in the structured generation action can be... The action is determined using the following formula (14). Corresponding attribution weights .

[0145] (14) In formula (14), Indicates action embedding or action type. This is the attribution scoring function.

[0146] Step S9: Determine the local reward corresponding to the action based on the reinforcement learning reward, attribution weight, and the corresponding score of the action in the evaluation results of the candidate video.

[0147] In this invention, the action is assigned to this action. Local rewards It can be calculated using the following formula (15).

[0148] (15) In this invention, after obtaining the local rewards corresponding to all actions in the structured generation action, the correspondence between actions and local rewards can be used to update the strategy of the video generation agent, making it clearer which actions are more important during the update process and accurately updating the strategy for generating the action.

[0149] The beneficial technical effects achieved by steps S8 to S9 are as follows: By utilizing the action dependency graph, the local reward assigned to each action in the structured generation action is calculated, so that different intelligent agents can obtain training signals that match their true contributions. The intelligent computing cloud platform can more accurately optimize different actions instead of letting all modules blindly respond to the same total score, thereby improving the efficiency of policy update.

[0150] For example, in a short drama generation scenario, the user simply inputs "create a short workplace drama clip with a strong plot twist." The video generation AI first generates the plot setting, character relationships, a three-part storyboard, a description of each shot, and video generation prompts. After candidate videos are generated, they are evaluated and scores are output based on plot completeness, conflict progression, plot twist effectiveness, character action plausibility, shot continuity, and image stability.

[0151] Because some candidate videos have high image quality but lack sufficient plot twists and causal closure at the end; or because other candidate videos have average image quality but good plot structure, the intelligent computing cloud platform can allocate plot-related rewards primarily to action. and The image quality and timing stability bonuses are primarily allocated to the action. , and Therefore, the strategy update is not simply to change all actions in the same overall direction, but to strengthen different capabilities such as "clear plot twists", "cameras serving the plot", "cue words supplementing the details of the main action", and "sampling parameters avoiding violent flickering".

[0152] As can be seen from this embodiment, the present invention does not simply feed the test scores back into the generative model, but decomposes the multi-dimensional test results and attributes them to the agent's action chain, thereby improving the learning effectiveness.

[0153] In one embodiment, prior to step S5, the method further includes: Step S10: Parse, clean, and structure the business rules to obtain a set of entries.

[0154] In this invention, an entry set can be obtained by parsing, cleaning, and structuring an external rule document containing business rules using a document adapter. The entry set can contain multiple entries as shown in the following formula (16): (16) In formula (16), each entry It includes at least one of the following: tag, description, tier, weight, hard threshold attribute, reward mapping method, and applicable scenario.

[0155] Step S11: Based on the set of entries, form the reward configuration for the reward compiler.

[0156] In this invention, the reward compiler is based on the set of entries. The reward configuration is automatically generated as shown in the following formula (17): (17) In formula (17), Represents the threshold set, This indicates cost and compliance constraints.

[0157] The beneficial technical effects achieved by steps S10 to S11 are that business rules are no longer just evaluation prompts, but reward rule assets that can directly drive reinforcement learning to evolve.

[0158] For example, in competition or customer acceptance scenarios, business users can upload a video scoring rules document, which includes requirements such as theme relevance, visual clarity, story completeness, creative expression, character consistency, duration limits, and prohibited elements. The document adapter can parse this document into structured Rubric entries and mark which entries are hard thresholds, which entries are weighted rewards, and which entries are safety penalties.

[0159] Then, the reward compiler can generate a reward configuration based on this. Subsequently, when the video generation agent generates videos under this business version, it not only pursues general video quality but also explicitly optimizes the topic relevance, story completeness, and ability to avoid disabled elements corresponding to the document. If the business side subsequently modifies the rules, the intelligent computing cloud platform can create new Rubric versions and reward configuration versions, and compare the differences between the old and new versions using a fixed regression set. The new version will only enter production after the quality benefits, cost benefits, and dispute rate all meet the conditions.

[0160] This embodiment illustrates that the present invention can elevate external rules from "evaluation criteria" to "self-evolving reward criteria" and ensure their controllable deployment through version governance.

[0161] In one embodiment, the structured generation action includes model selection, and step S3 includes: Step S3.1: Based on model selection, determine at least one model to be used for generating the video; Step S3.2: Call at least one model to generate candidate videos based on structured generation actions.

[0162] The beneficial technical effects achieved in steps S3.1 to S3.2 are as follows: By including a model selection dimension in the structured actions, the agent can call multiple models in parallel to generate candidate videos for the same user request. Reinforcement learning automatically learns "what kind of request should be assigned to which model" by comparing the gating rewards and group relative advantages of different models, enabling the self-evolving agent's policy to be bound to multiple models, improving the diversity of video generation, and achieving efficient utilization of computing power.

[0163] In one embodiment, the intelligent computing cloud platform implements a method for the self-evolution of video-generating intelligent agents through computing power, and before step S2, it further includes: Step S12: Based on the current task budget for the self-evolution of the video generation agent, determine the computing power orchestration strategy of the intelligent computing cloud platform.

[0164] The above computing power orchestration strategy can control at least one of the following: the number of candidate videos generated, generation parameters, evaluation dimensions, and training frequency of the video generation agent.

[0165] In one embodiment, the cost of reinforcement learning for video generation is significantly higher than that for text tasks. Therefore, this invention introduces a computing power orchestration strategy in an intelligent computing cloud platform to jointly control the number of candidates, resolution, frame rate, number of generation steps, evaluation depth, and training frequency. This allows for the scheduling of GPU, CPU, storage, or network resources for video generation, evaluation, training, caching, playback, and regression tasks.

[0166] In this invention, the cost required for the video generation agent to generate candidate videos can be calculated using the following formula (18).

[0167] (18) In formula (18), Indicates the production cost. Indicates the cost of the assessment. Indicates storage cost, This represents the training cost of an agent's self-evolution, i.e., the computational overhead of policy updates.

[0168] Given that the task budget for the self-evolution of the current video generation agent is determined to be B based on actual needs, the computational power orchestration strategy should control... That is, when the budget is insufficient, strategies such as low-resolution pre-screening, rapid keyframe evaluation, early candidate stopping, cache hits, and re-evaluation of a small number of high-confidence candidates are prioritized; when the budget is sufficient and the task value is high, the number of candidate videos generated and the depth of evaluation of candidate videos can be increased. Thus, reinforcement learning self-evolution is no longer an unconstrained consumption of computing power, but is executed within the budget of the cloud platform.

[0169] The beneficial technical effect achieved by step S12 is that, through computational power orchestration and adaptive candidate number, the reinforcement learning self-evolution process adapts to the resource constraints of the cloud platform production environment.

[0170] For example, for ordinary tasks, the intelligent computing cloud platform can generate fewer candidate videos and use rapid evaluation; for high-value tasks or tasks with a high historical failure rate, the intelligent computing cloud platform can automatically increase the number of candidate videos and the evaluation depth. The computing power orchestration module dynamically adjusts the number of candidate videos K, resolution, and evaluation level based on GPU queues, task priorities, historical hit cache, and budget limits.

[0171] When GPU resources are scarce, candidates can be quickly screened using low-resolution or short clips, followed by full generation and in-depth evaluation of high-potential candidates. When resources are plentiful, the intelligent computing cloud platform generates more candidates to form a more stable group relative advantage. This mechanism enables self-evolutionary training and production services to coexist on the same intelligent computing cloud platform.

[0172] It should be noted that, in some possible embodiments of the present invention, the method of enabling self-evolution of video generation agents through computing power on the intelligent computing cloud platform, in scenarios of private deployment or trainable base models, can, in addition to updating the agent policy, also train lightweight adapters, such as Low-Rank Adaptation (LoRA), control modules, style adapters, or post-processing networks. In this case, the reward obtained by encoding the evaluation results can simultaneously apply to the policy parameters. and adapter parameters The updated strategy parameters and adapter parameters are shown in the following formula (19): (19) In formula (19), To generate trajectory data, The reward set is defined. The intelligent computing cloud platform prevents adapter overfitting of the scorer through KL constraints, regression set testing, and canary releases. This embodiment is applicable to enterprise-owned video generation models or publicly available video generation models that can be fine-tuned locally.

[0173] As described above, this invention can compile multi-dimensional evaluation results into stable, interpretable, and manageable rewards; attribute rewards to different agent actions such as plot planning, storyboard generation, cue word expansion, model selection, sampling parameters, and post-processing; complete reinforcement learning updates with controllable computing power in expensive video generation scenarios; prevent agents from sacrificing real user experience by only pursuing evaluation scores; and achieve version rollback and policy auditability after rule changes.

[0174] Therefore, the intelligent computing cloud platform proposed in this invention, which enables the self-evolution of video generation agents through computing power, can solve problems such as the difficulty of demonstrating creativity through a closed-loop technology in simple video evaluation, the inability of a single reward model to cover complex business, the difficulty of attribution optimization for multi-step video generation agents, the excessive cost of video reinforcement learning, and the difficulty of governance after rule changes. Through this invention, the video generation system can evolve from a one-time generation tool into a production-grade intelligent agent system with sustainable learning, sustainable optimization, and sustainable governance, applicable to various scenarios such as short drama generation, advertising video generation, product demonstration video generation, character animation, image-generated video, audio-driven video, and enterprise private video generation platforms.

[0175] Please see Figure 2 , Figure 2 This is a structural diagram of a device provided by the present invention that enables the self-evolution of video-generating intelligent agents through computing power in an intelligent computing cloud platform. Figure 2 As shown, the device 200 for enabling the self-evolution of video-generating intelligent agents through computing power in an intelligent computing cloud platform includes: The acquisition module 201 is used to acquire user video generation requirements and corresponding business rules; The intelligent agent module 202 is used to input the user's video generation requirements into the video generation intelligent agent, schedule the computing power of the intelligent computing cloud platform, and obtain at least one set of structured generation actions output by the video generation intelligent agent. The video generation module 203 is used to schedule the computing power of the intelligent computing cloud platform to generate at least one candidate video based on at least one set of structured generation actions. Evaluation module 204 is used to perform multi-dimensional video evaluation on each candidate video based on business rules and obtain evaluation results; Reward compiler module 205 is used to transform the evaluation results into reinforcement learning rewards using the reward compiler; Update module 206 is used to update the strategy of the video generation agent based on the reinforcement learning rewards and structured generation actions corresponding to all candidate videos. The verification module 207 is used to perform version regression verification on the updated strategy of the video generation agent, and release the updated strategy when the verification results meet the release conditions, thus obtaining the self-evolved video generation agent.

[0176] In one embodiment, the evaluation module 204 can be specifically used for: Perform at least two of the following assessments on the candidate video: objective quality assessment, multimodal semantic assessment, text alignment assessment, originality assessment, rule adaptation assessment, cost assessment, and security assessment, to obtain the assessment vector; Determine the cost required to generate candidate videos; Based on the hard threshold rules included in the business rules, determine the hard threshold result corresponding to the candidate video. The hard threshold result includes parameter values ​​indicating whether the hard threshold rule is met, as well as the degree of violation of the hard threshold rule. The candidate videos are subjected to evidence consistency verification to obtain evidence consistency scores; Based on user preference data, determine the preference gain corresponding to the candidate video; The evaluation results are derived from the evaluation vector, cost, hard threshold results, evidence consistency score, and preference gain.

[0177] In one embodiment, the reward compilation module 205 can be used specifically for: Based on the weight of each component in the evaluation vector in the business rules, all components are weighted and summed to obtain the basic reward. Based on the evidence consistency score, an abnormal speculation risk value is determined, which is used to describe the penalty index in the reward construction of the video generation agent's reinforcement learning. The reinforcement learning reward is obtained by integrating the basic reward with the abnormal speculative risk value, cost, degree value, and preference gain.

[0178] In one embodiment, the reward compilation module 205 can also be used for: Based on the parameter values ​​in the hard threshold results, reinforcement learning rewards, and preset failure penalties, the gating reward for candidate videos is determined, and the gating reward is used to update the policy of the video generation agent.

[0179] In one embodiment, the update module 206 can be specifically used for: Based on the gating rewards corresponding to all candidate videos, determine the relative group advantage for each candidate video; The strategy of the video generation agent is updated based on the relative advantages of the groups corresponding to the candidate videos and the structured generation actions.

[0180] In one embodiment, the update module 206 can be specifically used for: For any two candidate videos, construct preference samples based on the reinforcement learning rewards and structured generation actions corresponding to the two candidate videos respectively; Update the policy of the video generation agent based on the preference samples.

[0181] In one embodiment, a reward attribution module may also be included for: Based on a pre-built action dependency graph, the attribution weight of each action in the structured generated action is determined, where the action dependency graph contains the mapping relationship between multiple actions and video evaluation metrics. The local reward for an action is determined based on the reinforcement learning reward, attribution weight, and the corresponding score of the action in the evaluation results for the candidate video.

[0182] In one embodiment, a document adaptation module may also be included, for: The business rules are parsed, cleaned, and structured to obtain a set of entries. The set of entries contains multiple entries, and each entry contains at least one of the following: tag, description, tier, weight, hard threshold attribute, reward mapping method, and applicable scenario. Based on the set of entries, the reward configuration for the reward compiler is formed.

[0183] In one embodiment, the video generation module 203 can be specifically used for: Based on model selection, determine at least one model to be used for video generation; Invoke at least one model to generate candidate videos based on structured generation actions.

[0184] In one embodiment, it may further include a computing power orchestration module, used for: Based on the current task budget for the self-evolution of the video generation agent, a computing power orchestration strategy for the intelligent computing cloud platform is determined. The computing power orchestration strategy controls at least one of the following: the number of candidate videos generated, generation parameters, evaluation dimensions, and training frequency of the video generation agent.

[0185] The device for enabling the self-evolution of video-generating intelligent agents through computing power provided by the intelligent computing cloud platform of the present invention is capable of realizing the various processes of the various embodiments of the above-mentioned method for enabling the self-evolution of video-generating intelligent agents through computing power of the intelligent computing cloud platform. The technical features are one-to-one and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0186] It should be noted that the intelligent computing cloud platform in this invention can be a device that enables the self-evolution of video-generated intelligent agents through computing power, or it can be a component, integrated circuit, or chip in an electronic device.

[0187] The present invention also provides an electronic device, see [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. The electronic device includes a memory 301, a processor 302, and a program or instructions stored in the memory 301 that run on the memory. When the program or instructions are executed by the processor 302, they can achieve the following: Figure 1 The corresponding intelligent computing cloud platform achieves any step in the method embodiment of video generation intelligent agent self-evolution through computing power and achieves the same beneficial effect, which will not be elaborated here.

[0188] The processor 302 can be a CPU, ASIC, FPGA or GPU.

[0189] Those skilled in the art will understand that all or part of the steps of the above-described intelligent computing cloud platform for achieving self-evolution of video-generating intelligent agents through computing power can be implemented by hardware related to program instructions, and the program can be stored in a readable medium.

[0190] The present invention also provides a readable storage medium on which a computer program is stored, and which, when executed by a processor, can perform the above-described functions. Figure 1 The corresponding intelligent computing cloud platform can achieve any step in the method embodiment of the self-evolution of video-generating intelligent agents through computing power, and can achieve the same technical effect. To avoid repetition, it will not be described again here. The storage medium mentioned is such as read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc.

[0191] The present invention also provides a computer program product, including computer instructions that, when executed by a processor, implement the above-described... Figure 1 The corresponding intelligent computing cloud platform realizes each process of the self-evolution method embodiment of video generation intelligent agent through computing power, and can achieve the same technical effect. To avoid repetition, it will not be described in detail here.

[0192] The terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses. Additionally, the use of "and / or" in this application indicates at least one of the connected objects, such as A and / or B and / or C, representing seven possibilities: A alone, B alone, C alone, both A and B present, both B and C present, both A and C present, and A, B, and C present.

[0193] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0194] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or second terminal device, etc.) to execute the methods of the various embodiments of this application.

[0195] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for achieving self-evolution of video-generating intelligent agents through computing power on an intelligent computing cloud platform, characterized in that, include: Step S1: Obtain user video generation requirements and corresponding business rules; Step S2: Input the user's video generation requirements into the video generation agent, schedule the computing power of the intelligent computing cloud platform, and obtain at least one set of structured generation actions output by the video generation agent. The structured generation actions include at least one of the following: plot planning, storyboard planning, prompt words, model selection, generation parameters, and post-processing parameters. Step S3: Based on the at least one set of structured generation actions, schedule the computing power of the intelligent computing cloud platform to generate at least one candidate video; Step S4: Based on the business rules, perform multi-dimensional video evaluation on each candidate video to obtain the evaluation results; Step S5: Using a reward compiler, the evaluation results are converted into reinforcement learning rewards; The method further includes, after step S5: determining the attribution weight of each action in the structured generated action corresponding to the candidate video based on a pre-constructed action dependency graph, wherein the action dependency graph contains multiple mapping relationships between actions and video evaluation metrics, and the existence of a mapping relationship between actions and video evaluation metrics indicates that the magnitude of the video evaluation metric in the candidate video score is affected by the action. The local reward corresponding to the action is determined based on the reinforcement learning reward of the candidate video, the attribution weight, and the score of the action in the evaluation result. Step S6: After obtaining the local reward, the correspondence between the action and the local reward is used to update the strategy of the video generation agent; Step S7: Perform version regression verification on the updated strategy of the video generation agent, and release the updated strategy when the verification result meets the release conditions to obtain the self-evolved video generation agent.

2. The method as described in claim 1, characterized in that, Step S4 includes: Step S4.1: Perform at least two of the following on the candidate video: objective quality evaluation, multimodal semantic evaluation, text alignment evaluation, originality evaluation, rule adaptation evaluation, cost evaluation, and security evaluation, to obtain the evaluation vector; Step S4.2: Determine the cost required to generate the candidate video; Step S4.3: Based on the hard threshold class rules included in the business rules, determine the hard threshold result corresponding to the candidate video. The hard threshold result includes a parameter value indicating whether the hard threshold class rule is satisfied, and a value indicating the degree of violation of the hard threshold class rule. Step S4.4: Perform evidence consistency verification on the candidate videos to obtain evidence consistency scores; Step S4.5: Based on the user's preference data, determine the preference gain corresponding to the candidate video; Step S4.6: Obtain the evaluation result from the evaluation vector, the cost, the hard threshold result, the evidence consistency score, and the preference gain.

3. The method as described in claim 2, characterized in that, Step S5 includes: Step S5.1: Based on the weight of each component in the evaluation vector in the business rules, perform a weighted summation of all components to obtain the basic reward; Step S5.2: Based on the evidence consistency score, determine the abnormal speculation risk value, which is used to describe the penalty index in the reward construction of the video generation agent reinforcement learning. Step S5.3: Combine the basic reward with the abnormal speculative risk value, the cost, the degree value, and the preference gain to calculate the reinforcement learning reward.

4. The method as described in claim 3, characterized in that, Following step S5.3, step S5 further includes: Step S5.4: Based on the parameter values ​​in the hard threshold result, the reinforcement learning reward, and the preset failure penalty, determine the gating reward for the candidate video, and use the gating reward to update the strategy of the video generation agent.

5. The method as described in claim 4, characterized in that, Step S6 includes: Step S6.1: Determine the relative advantage of each candidate video group based on the gating rewards corresponding to all candidate videos. Step S6.2: Update the strategy of the video generation agent based on the relative advantages of the groups corresponding to the candidate videos and the structured generation actions.

6. The method as described in claim 1, characterized in that, Step S6 includes: Step S6.1': For any two candidate videos, construct preference samples based on the reinforcement learning rewards and structured generation actions corresponding to the two candidate videos respectively; Step S6.2': Update the strategy of the video generation agent based on the preference sample.

7. The method as described in claim 1, characterized in that, Before step S5, the following is also included: Step S10: Parse, clean and structure the business rules to obtain an item set. The item set contains multiple items, and each item contains at least one of the following: tag, description, tier, weight, hard threshold attribute, reward mapping method and applicable scenario. Step S11: Based on the set of entries, form the reward configuration of the reward compiler.

8. The method as described in claim 1, characterized in that, The structured generation action includes model selection, and step S3 includes: Step S3.1: Based on the model selection, determine at least one model for generating the video; Step S3.2: Invoke the at least one model and generate candidate videos based on the structured generation action.

9. The method as described in claim 1, characterized in that, Before step S2, the following is also included: Step S12: Based on the current task budget of the self-evolution of the video generation agent, determine the computing power orchestration strategy of the intelligent computing cloud platform. The computing power orchestration strategy controls at least one of the following: the number of candidate videos generated, generation parameters, evaluation dimensions, and training frequency of the video generation agent.

10. A device for enabling self-evolution of video-generating intelligent agents through computing power in an intelligent computing cloud platform, characterized in that, include: The acquisition module is used to acquire user video generation requirements and corresponding business rules; The intelligent agent module is used to input the user's video generation requirements into the video generation intelligent agent, schedule the computing power of the intelligent computing cloud platform, and obtain at least one set of structured generation actions output by the video generation intelligent agent. The structured generation actions include at least one of the following: plot planning, storyboard planning, prompt words, model selection, generation parameters, and post-processing parameters. The video generation module is used to schedule the computing power of the intelligent computing cloud platform to generate at least one candidate video based on the at least one set of structured generation actions. The evaluation module is used to perform multi-dimensional video evaluation on each candidate video based on the business rules and obtain the evaluation results; The reward compilation module is used to transform the evaluation results into reinforcement learning rewards using a reward compiler. The reward compilation module, after converting the evaluation results into reinforcement learning rewards, is further used for: Based on a pre-constructed action dependency graph, the attribution weight of each action in the structured generated actions corresponding to the candidate video is determined. The action dependency graph contains multiple mapping relationships between actions and video evaluation metrics. The existence of a mapping relationship between actions and video evaluation metrics indicates that the magnitude of the video evaluation metric in the candidate video score is affected by the action. The local reward corresponding to the action is determined based on the reinforcement learning reward of the candidate video, the attribution weight, and the score of the action in the evaluation result. An update module is used to update the strategy of the video generation agent by establishing the correspondence between the action and the local reward after obtaining the local reward. The verification module is used to perform version regression verification on the updated strategy of the video generation agent, and release the updated strategy when the verification result meets the release conditions, so as to obtain the self-evolved video generation agent.

11. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein when executed by the processor, the program implements the steps of the method for achieving self-evolution of a video-generating intelligent agent through computing power using the intelligent computing cloud platform as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for achieving self-evolution of video-generating intelligent agents through computing power using the intelligent computing cloud platform as described in any one of claims 1 to 9.

13. A computer program product, characterized in that, The method includes computer instructions that, when executed by a processor, implement the steps of the method for achieving self-evolution of a video-generating intelligent agent through computing power using the intelligent computing cloud platform as described in any one of claims 1 to 9.

Images