AI-based methods and systems for generating and publishing viral videos

Abstract

This invention discloses an AI-based method and system for generating and publishing viral videos. It realizes a dynamic memory network that, through a continuous time update mechanism and causal maintenance verification, ensures that the recombined content aligns with the surface intent of the user's marketing copy. Through the RAG+MAB architecture, it achieves fully autonomous decision-making, solving the problems of traditional publishing processes that require manual selection of platforms, timing, and tags, which are time-consuming and dependent on experience. Through the synergy of inner-loop content optimization and outer-loop publishing strategies, it addresses the shortcomings of generating good content but publishing it at the wrong time or optimizing the publishing strategy but resulting in insufficient content quality.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an AI-based method and system for generating and publishing viral videos. Background Technology

[0002] With the rapid development of mobile internet and short video platforms, short videos have become a core carrier of information dissemination and commercial marketing. How to efficiently and in batches generate personalized marketing videos with "viral" potential is a pressing technical challenge in the content creation field. Existing video generation technologies mainly include: template-based splicing methods, text-to-video methods based on generative adversarial networks or diffusion models, and personalized recommendation and generation methods based on deep learning. However, existing technologies lack a deep understanding of the viral mechanism, typically directly transferring style or content from the video as a whole, failing to decompose successful viral videos into quantifiable and reusable minimum creative units, and even less establishing a complex causal relationship graph between elements. This makes it difficult for generated content to inherit the deep-seated propagation dynamics of viral videos. Most existing systems rely on static training data, unable to absorb the trend characteristics of the latest viral videos in real time and dynamically update their knowledge structure, resulting in generated content that easily becomes outdated and disconnected from real-time hot topics. Traditional diffusion models, when generating videos, primarily focus on visual coherence and textual conditional alignment, lacking constraints on causal consistency between frame sequences and elements. This results in videos with logical inconsistencies, weak persuasiveness, and difficulty in replicating the effects of proven element combinations that go beyond simple superposition. In existing technologies, content generation and distribution strategies are typically optimized independently, failing to form an end-to-end closed-loop feedback system. The actual dissemination effect of the generated videos is insufficient to feed back into the generation model, nor can it adaptively adjust distribution decisions based on real-time market feedback. Summary of the Invention

[0003] This invention provides an AI-based method for generating and publishing viral videos, the method comprising the following steps:

[0004] S1. Decompose the monitored viral videos into atomized genes, which are the smallest reusable creative units, and establish a causal relationship graph between elements;

[0005] S2. Construct a real-time evolving knowledge base for trending videos, encapsulate each atomized gene as a multimodal memory unit, and dynamically update the node feature matrix and temporal adjacency matrix based on new trending video data through a temporal dynamic graph neural network update mechanism;

[0006] S3. Receive user marketing copy, generate dynamic query vectors through large model analysis, perform cross-modal retrieval in the dynamic multimodal memory network, find multiple most similar multimodal memory units, generate multimodal context vectors through attention-weighted aggregation, and perform causal preservation verification at the same time;

[0007] S4. Decompose the multimodal context vector into specific generation conditions, input them into a video generator based on a diffusion model, and generate video segments that conform to the causal structure by introducing causal consistency loss and temporal coherence loss.

[0008] S5. Optimization is achieved through a dual-loop reinforcement learning architecture: the inner loop models the video generation process as a Markov decision process and optimizes the generation model parameters using the policy gradient method; the outer loop optimizes the release decision model based on the proximal policy to optimize the release policy.

[0009] S6. Extract feature tags from the generated video, retrieve the optimal release parameters in the release strategy knowledge base by using the enhanced RAG retrieval method, and perform autonomous release; collect actual business data after release, and perform closed-loop updates on the time-series dynamic multimodal memory network, diffusion model, and release strategy model.

[0010] Furthermore, in step S1, the decomposition specifically includes: using a scene detection algorithm to analyze viral videos. Divided into K storyboard-level units Extract visual feature vectors for each scene. Text feature vectors Motion feature vectors And associate performance tags ; Construct a structural causal model ,in It is a set of atomized gene nodes. For causal relationship, The causal strength weight matrix is ​​used, and the causal emergence effect score for each atomized gene combination is calculated. .

[0011] Furthermore, in step S1, the causal emergence effect score is calculated. The calculation method is as follows:

[0012]

[0013] in, For elements arrive The strength of causal influence For the synergistic effect function, This is an emergent correction term.

[0014] Furthermore, in step S2, the temporal dynamic graph neural network update mechanism is specifically as follows:

[0015]

[0016]

[0017] in, The element feature matrix at time t+1 is the visual feature vector + text feature vector + motion feature vector. Let be the characteristic matrix of elements at time t; Let be the temporal adjacency matrix at time t, representing the dynamic correlation strength between popular elements; Let be the node attribute matrix at time t, containing the performance labels of the elements; These are learnable parameters; For graph neural networks; The temporal residual weight is used to control the proportion of historical memory retention. The map update rate controls the speed at which new knowledge is absorbed. Let be the temporal adjacency matrix at time t+1; Let be the query matrix at time t; The key matrix at time t+1; This is the value matrix at time t+1; For attention head dimension.

[0018] Furthermore, in step S3, generating the dynamic query vector specifically includes the following calculations:

[0019]

[0020] Parameter Q is the base query vector; Embedding functions for large language models; Marketing copy text entered by the user; User profile vectors; This is the intent recognition vector; This is a vector concatenation function;

[0021]

[0022] parameter The position-aware query vector for the j-th scene; Based on the query projection matrix; This is a position-encoded vector; The position is encoded as a projection matrix; The set of query vectors mentioned above represents the set of generated storyboard query vectors; This is the encoding function for the preceding context; This is the context projection matrix.

[0023] Further, in step S2, cross-modal retrieval is performed in the dynamic multimodal memory network to find multiple most similar multimodal memory units, wherein the cross-modal retrieval similarity scoring formula is:

[0024]

[0025] in, query vector With memory unit The overall similarity score; The cosine similarity function; For the r-th candidate memory unit, it is one of the Top-k candidates retrieved from the dynamic memory network; The text projection matrix of the position-aware query vector of the j-th scene; The visual projection matrix of the position-aware query vector for the j-th scene; The motion projection matrix for the position-aware query vector of the j-th scene; Let be the text feature vector of the r-th candidate memory unit; Let be the visual feature vector of the r-th candidate memory unit; Let r be the motion feature vector of the r-th candidate memory unit; These are the modal weighting coefficients.

[0026] Furthermore, in step S5, the formula for calculating the policy gradient of the inner loop optimization is:

[0027]

[0028] in, This represents the gradient of the policy objective function; Let the policy objective function be... To generate model parameters; This refers to the state-action sequence trajectory during the video generation process. To generate a strategy, determine the action to take in a given state; Calculate the expectation of all possible trajectories according to the policy distribution; For time steps; The state at time t; The action taken at time t; Estimate the content advantage function; The final generated video; These are the weighting coefficients for the prediction model; Scoring of the automatic prediction model; Baseline value; The score is determined by human evaluation; the content advantage function is... Score assessed by humans With automatic prediction model scoring The weighted combination is obtained by subtracting the baseline value.

[0029] This invention also relates to an AI-based viral video generation and publishing system, which uses the aforementioned AI-based viral video generation and publishing method, and the system includes:

[0030] The causal deconstruction and extraction module is used to decompose the monitored viral videos into atomized genes, which are the smallest reusable creative units, and to establish a causal relationship graph between elements.

[0031] The dynamic memory network module is used to build a real-time evolving knowledge base of trending videos. It encapsulates each atomized gene into a multimodal memory unit and updates the node feature matrix and temporal adjacency matrix dynamically based on new trending video data through a temporal dynamic graph neural network update mechanism.

[0032] The intelligent reorganization module is used to receive user marketing copy, generate dynamic query vectors through large model analysis, perform cross-modal retrieval in the dynamic multimodal memory network, find multiple most similar multimodal memory units, and generate multimodal context vectors through attention-weighted aggregation, while performing causal preservation verification.

[0033] The causal constraint generation module is used to decompose the multimodal context vector into specific generation conditions, input them into the video generator based on the diffusion model, and generate video segments that conform to the causal structure by introducing causal consistency loss and temporal coherence loss.

[0034] The dual-loop reinforcement learning module is used for optimization through a dual-loop reinforcement learning architecture: the inner loop models the video generation process as a Markov decision process and optimizes the generation model parameters using the policy gradient method; the outer loop optimizes the release decision model based on the proximal policy to optimize the release policy.

[0035] The autonomous release and evolution module is used to extract feature tags from the generated video, retrieve the optimal release parameters in the release strategy knowledge base by retrieving enhanced RAGs, and execute autonomous release; collect actual business data after release, and perform closed-loop updates on the temporal dynamic multimodal memory network, diffusion model, and release strategy model.

[0036] The present invention also relates to a computer-readable storage medium for storing a computer program, which is executed by a processor for performing the above-described AI-based method for generating and publishing viral videos.

[0037] The present invention also relates to a computer program product, the computer program product comprising a computer program, which is executed by a processor for performing the above-described AI-based method for generating and publishing viral videos.

[0038] This invention implements a dynamic memory network that, through a continuous time update mechanism, enables the system to capture emerging trends such as prioritizing emotional value and reversing narratives in real time. Simultaneously, causal preservation verification ensures that the reorganized content aligns with the surface-level intent of user marketing copy. Furthermore, the RAG+MAB architecture of this invention achieves fully autonomous decision-making, solving the problems of traditional publishing processes requiring manual selection of platforms, timing, and tags, which is time-consuming and reliant on experience. The synergy between inner-loop content optimization and outer-loop publishing strategies addresses the shortcomings of generating good content but publishing it at the wrong time, or optimizing the publishing strategy but resulting in insufficient content quality. Attached Figure Description

[0039] Figure 1 This is a flowchart of the AI-based method for generating and publishing viral videos according to the present invention. Detailed Implementation

[0040] The present invention will now be further described with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention. It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention.

[0041] Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular forms are intended to include the plural forms as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] See appendix Figure 1 Embodiment 1 of the present invention relates to an AI-based method for generating and publishing viral videos, comprising the following steps:

[0043] S1. The monitored viral videos are decomposed into atomized genes, which are the smallest reusable creative units, and a causal relationship graph between elements is established to provide causal mechanism support for subsequent intelligent recombination.

[0044] Video segmentation and multimodal feature extraction include using scene detection algorithms to identify viral videos. Divided into Each storyboard-level unit Multi-dimensional feature extraction is performed on each scene.

[0045] Visual feature extraction:

[0046]

[0047] in, Indicates the first One viral video ( ) Each storyboard ( The visual feature vector of ); For CLIP vision encoder; This is the keyframe extraction function; For video The One storyboard; is the dimension of the visual feature vector.

[0048] Text feature extraction:

[0049] The dialogue text is extracted through automatic speech recognition, analyzed by a large model to generate narrative structure tags and emotional trigger point tags, and encoded into text feature vectors:

[0050]

[0051] in, Indicates the first One viral video ( ) Each storyboard ( The text feature vector of ); It is a multilayer perceptron fusion network; The dimension of the text feature vector; For text semantic vectors; Embed vectors for narrative structure labels; Embed vectors for sentiment tags.

[0052] Motion feature extraction:

[0053]

[0054] in, Indicates the first One viral video ( ) Each storyboard ( The motion feature vector; It is a three-dimensional convolutional neural network; Let be the dimension of the motion feature vector; This is an algorithm for calculating pixel-level motion vectors in an image sequence.

[0055] Performance tag association:

[0056] Link the video to historical business metrics

[0057]

[0058] in, This is the performance tag for this storyboard; For the first One viral video ( ) Each storyboard ( The completion rate of the broadcast; For the first One viral video ( ) Each storyboard ( Conversion rate; For the first One viral video ( ) Each storyboard ( Click-through rate; For the first One viral video ( ) Each storyboard ( ) sharing rate.

[0059] Structural causal model construction and causal strength calculation, including constructing structural causal models of trending elements and identifying causal relationships between elements:

[0060]

[0061] in, E represents the set of all extracted atomized gene nodes; W represents the causal relationships learned from massive amounts of viral video data using a causal algorithm; G represents the weight matrix for calculating causal strength; and G represents the structural causal model diagram of viral elements.

[0062] The elements of the weight matrix W , represented as element right The strength of causal influence:

[0063]

[0064] Indicates a pair of elements Perform the intervention operation and set it to x; Represents element The mean; For expectation operators.

[0065] The causal emergence effect score includes calculating the causal emergence effect score for each atomized gene combination.

[0066]

[0067]

[0068]

[0069] in, Represents element pairs The causal emergence effect score; This represents the synergistic effect function, which measures the additional effect produced by the combination of elements. To capture nonlinear interaction effects as emergent correction terms; is a learnable collaborative matrix; d is the feature dimension of the atomized gene element; It is the sigmoid activation function; Mutual information measures the statistical dependency between elements; Information entropy; The feature vector representing the combination of elements; This is the balance coefficient.

[0070] S2. Construct a popular knowledge base that can evolve in real time, encapsulating each atomized gene into a multimodal memory unit. Furthermore, it uses a temporal dynamic graph neural network update mechanism to dynamically update the node feature matrix based on new trending video data. Temporal adjacency matrix .

[0071] By constructing a real-time evolving knowledge base of trending topics, storing multimodal features of atomized genes, and using dynamic graph neural networks to achieve temporal updates of knowledge, the system ensures it grasps the latest trending topics. Specifically, this includes:

[0072] Multimodal memory unit encapsulation. Each atomized gene is encapsulated as a multimodal memory unit.

[0073]

[0074] in, For the first One viral video ( ) Each storyboard ( () multimodal memory units; For the causal context, it includes the causal influence weight and emergent effect score of the element; This is a timestamp, recording the time when the element entered the knowledge base; For visual feature vectors; For text feature vectors; For motion feature vectors; The performance tag for this element.

[0075] Temporal dynamic graph neural network update. Constructing a knowledge evolution mechanism based on dynamic graph neural networks:

[0076]

[0077]

[0078] in, The element feature matrix at time t+1 is the visual feature vector + text feature vector + motion feature vector. Let be the characteristic matrix of elements at time t; Let be the temporal adjacency matrix at time t, representing the dynamic correlation strength between popular elements; Let be the node attribute matrix at time t, containing the performance labels of the elements; These are learnable parameters; For graph neural networks; The temporal residual weight is used to control the proportion of historical memory retention. The map update rate controls the speed at which new knowledge is absorbed. Let be the temporal adjacency matrix at time t+1; Let be the query matrix at time t; The key matrix at time t+1; This is the value matrix at time t+1; For attention head dimension.

[0079] S3. Receive user marketing copy and generate dynamic query vectors through large-scale model parsing. Cross-modal retrieval is performed in the dynamic multimodal memory network to find the top-k most similar multimodal memory units. Multimodal context vectors are generated through attention-weighted aggregation. At the same time, causal preservation verification is performed.

[0080] For user-specific marketing copy, the system retrieves the most relevant trending elements from a dynamic memory network and intelligently reorganizes them through semantic-causal joint optimization to ensure that the generated content aligns with marketing intent and possesses a viral dissemination mechanism. Specifically, this includes:

[0081] User intent parsing and dynamic query construction. Receives user marketing copy and generates dynamic query vectors through large-scale model analysis.

[0082]

[0083] Where Q is the basic query vector; Embedding functions for large language models; Marketing copy text entered by the user; User profile vectors; This is the intent recognition vector; This is a vector concatenation function.

[0084]

[0085] in, The position-aware query vector for the j-th scene; Based on the query projection matrix; This is a position-encoded vector; The position is encoded as a projection matrix; The set of query vectors mentioned above represents the set of generated storyboard query vectors; This is the encoding function for the preceding context; This is the context projection matrix.

[0086] Cross-modal retrieval and attention-weighted aggregation. Perform cross-modal retrieval in a dynamic memory network to find the top-k most similar multimodal memory units. ),

[0087]

[0088] in, query vector With memory unit The overall similarity score; The cosine similarity function; For the r-th candidate memory unit, it is one of the Top-k candidates retrieved from the dynamic memory network; The text projection matrix of the position-aware query vector of the j-th scene; The visual projection matrix of the position-aware query vector for the j-th scene; The motion projection matrix for the position-aware query vector of the j-th scene; Let be the text feature vector of the r-th candidate memory unit; Let be the visual feature vector of the r-th candidate memory unit; Let r be the motion feature vector of the r-th candidate memory unit; These are the modal weighting coefficients.

[0089]

[0090] in, Let r be the attention weight of the r-th memory unit; This represents the number of the top-k most similar multimodal memory units found. Indicates the combined performance-aging amplification factor; A performance data set; For memory unit timestamps; It is an exponential function.

[0091]

[0092] in, Let be the multimodal context vector of the j-th scene; Let r be the attention weight of the r-th memory unit; Let be the text feature vector of the r-th candidate memory unit; Let be the visual feature vector of the r-th candidate memory unit; Let r be the motion feature vector of the r-th candidate memory unit; Let r be the causal context vector of the r-th unit; This represents the number of the top-k most similar multimodal memory units found. This is a vector concatenation function.

[0093] Causality Preservation Verification. To ensure that the recombined element sequence maintains the validity of the causal structure, a causal preservation score is calculated. If the score is less than a threshold, a re-retrieval or insertion of transitional elements is triggered.

[0094] S4. Transfer the multimodal context vector Decomposed into specific generation conditions The input is fed into a diffusion-based video generator, and causal consistency loss is introduced. and temporal coherence loss , so that the total loss function This generates video clips that conform to a causal structure. .

[0095] Based on the recombined multimodal context vectors, personalized videos with coherent visuals and viral potential are generated, and causal constraints ensure that the generated content inherits verified propagation mechanisms. This includes...

[0096] Multi-condition controlled video generation. Multimodal context vectors... Break it down into specific generation conditions:

[0097]

[0098] in, The specific conditions for generating the decomposition are visual conditions. Text conditions Structural conditions A set of.

[0099] The input is fed into a diffusion-based video generator, which uses a denoising diffusion probability model to generate the video.

[0100] Causal consistency constraint. A causal consistency constraint is introduced during the training of the diffusion model to ensure that the generated video maintains consistency with the causal structure of the template.

[0101]

[0102] in, This is the total loss function of the causal constraint diffusion model; The loss is for the standard denoising diffusion model; Loss of causal consistency; This represents a loss of temporal continuity. This is the coefficient of causal constraint strength. This represents the temporal constraint strength coefficient.

[0103] Video assembly and post-optimization. This involves assembling and optimizing the video clips generated from each scene. The videos are assembled in sequence according to their narrative function to form a complete video. Quality enhancement is achieved through post-video optimization.

[0104] S5. Optimization is achieved through a dual-loop reinforcement learning architecture: the inner loop models the video generation process as a Markov decision process and optimizes the generation model parameters using the policy gradient method; the outer loop constructs a release decision model based on proximal policy optimization (PPO) to optimize the release strategy.

[0105] A dual-loop reinforcement learning architecture is used to collaboratively optimize content generation and publishing strategies: the inner loop optimizes video content quality, and the outer loop optimizes publishing decisions, forming an end-to-end performance improvement closed loop. Specifically, this includes...

[0106] Inner-loop optimization - content quality enhancement. The generation process in step S4 is modeled as a Markov decision process, and the generation parameters are optimized using the policy gradient method.

[0107]

[0108]

[0109] in, This represents the gradient of the policy objective function; Let the policy objective function be... To generate model parameters; This refers to the state-action sequence trajectory during the video generation process. To generate a strategy, determine the action to take in a given state; Calculate the expectation of all possible trajectories according to the policy distribution; For time steps; The state at time t; The action taken at time t; Estimate the content advantage function; The final generated video; These are the weighting coefficients for the prediction model; Scoring of the automatic prediction model; Baseline value; The score is determined by human evaluation.

[0110] Outer-loop optimization - release strategy reinforcement. Constructing a release decision model based on near-end strategy optimization:

[0111]

[0112]

[0113] in, The objective of the loop strategy optimization is to prune the loss function for PPO. To publish policy network parameters; Let be the expectation with respect to time step t; For strategy ratio; To publish the strategy advantage function; This is the clipping function; For clipping hyperparameters; For immediate reward functions, it represents the action to be performed. Immediate feedback on business metrics was obtained afterward; The state at time t; The action taken at time t; For RAG retrieval weight coefficients; Estimating the future value of RAG enhancement; This represents the state at time t+1; Publish a historical case memory bank; This is for estimating the state value function.

[0114] S6. Extract feature tags from the generated video, retrieve the optimal release parameters in the release strategy knowledge base by using the enhanced RAG retrieval method, and perform autonomous release; collect actual business data after release, and perform closed-loop updates on the time-series dynamic multimodal memory network, diffusion model, and release strategy model.

[0115] This process enables fully autonomous video publishing, including platform selection, timing optimization, and tag recommendations, and allows for continuous self-evolution of the system through post-publishing data feedback. Specifically, this includes...

[0116] RAG-enhanced publishing strategy retrieval. Extracting feature tags from generated videos. The RAG technology is used to retrieve the most strategic deployment strategy from the deployment strategy knowledge base and output the optimal deployment parameters. ( )

[0117]

[0118] in, Large language models enhanced for retrieval; ; For real-time information on the platform; This is a Top-K retrieval function that retrieves the most similar historical posting examples from a knowledge base. This is a similarity calculation function; Embed vectors for video features; Embed the matrix into the knowledge base.

[0119] Multi-armed slot machine platform selection. Platform selection is based on a multi-armed slot machine model, incorporating RAG knowledge information gain.

[0120]

[0121] in, To select the optimal platform; This indicates the value of the parameter to be maximized; A collection of optional platforms; The average experience reward for platform a; t is the current time step; The number of times platform a was selected; This refers to the knowledge gain weighting coefficient. For RAG knowledge information gain.

[0122] Closed-loop feedback and system self-evolution.

[0123] After the video is released, collect actual business data. Perform a system update.

[0124] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below with reference to specific embodiments. This embodiment takes an intelligent marketing short video generation system for e-commerce platforms as an example to illustrate the specific implementation process of this invention.

[0125] S1: Causal Emergence Deconstruction and Atomized Gene Extraction of Viral Videos. The system first connects to the APIs of short video platforms such as Douyin and Kuaishou, monitoring in real time and acquiring metrics such as play count, completion rate, and like rate that exceed preset thresholds, for example, the top 5% of viral videos. Let's assume we acquire a viral video... For example, an advertisement for a skincare product.

[0126] Video segmentation: Using the content-change-based scene detection algorithm from the PySceneDetect library, the video is segmented... Divided into 15 scenes arrive .For example, A close-up of "problem skin" For "the moment of product application", This is a comparison of skin radiance after use.

[0127] Multimodal feature extraction, including,

[0128] Visual features: Keyframes are extracted for each scene, input into a pre-trained CLIP visual encoder (ViT-B / 32), and output a 512-dimensional visual feature vector. .

[0129] Text features: Whisper model was used to extract dialogue from storyboards, and GPT-4 model was used for analysis. For example, for... The dialogue "Look, this essence is absorbed in a second" generates narrative structure tags "product efficacy demonstration" and emotional trigger tags "surprise / efficacy," which are then fused into a 256-dimensional text feature vector through a two-layer MLP. .

[0130] Motion features: Calculate the optical flow field of the storyboard, input it into a lightweight 3D-CNN (such as an I3D network), and extract a 128-dimensional motion feature vector. .For example, The "smearing" action in the text is characterized by rapid, rotating motion.

[0131] Performance tags: These are historical business metrics associated with the video. For example, completion rate Conversion rate Click-through rate sharing rate .

[0132] Structural causal model construction: The system has learned a structural causal model diagram from a massive amount of historical viral videos. In this example, the identified causal relationships include: "problem presentation" ( → “Product Introduction” → "Efficacy Verification" ). Causality strength The calculation is the rate of change in the expected probability of the "product introduction" element when the intensity of the "problem display" element is intervened.

[0133] Causal Emergent Effect Score: Calculate element pairs The score. Synergistic effect function. Through a learnable collaborative matrix Calculate the additional effects produced by combining the two (such as the synergistic effect of "quick application" + "gloss contrast" creating a sense of "highly effective improvement"). Emergent corrections. Calculate the mutual information and combinatorial entropy between the two. Final score. The high value indicates that the combination of "efficacy demonstration" and "effect comparison" can produce an emergent effect of 1+1>2.

[0134] S2. Construction of a temporal dynamic multimodal memory network.

[0135] Multimodal memory unit encapsulation: Encapsulating the features of each scene into a multimodal memory unit. Stored in the knowledge base.

[0136] Time-series dynamic graph neural network update: Assuming a new viral video emerges the next day, containing a new segment on "blue copper peptide ingredient science popularization." The system executes an update:

[0137] Initial characteristics of the new node This is determined by its multimodal characteristics.

[0138] Update the node feature matrix using dynamic GNN ,in It preserved the historical memory of the previous day's "ingredient display".

[0139] Temporal adjacency matrix In a synchronized update, a new strong correlation edge will be established between "Blue Copper Peptide Science Popularization" and "Troubleshooting," while the existing correlation weight of "Salicylic Acid Science Popularization" will be adjusted based on a time-series decay factor. By appropriately reducing the threshold, the dynamic evolution of knowledge is achieved.

[0140] S3. Intelligent recombination based on semantic-causal joint matching. User inputs marketing copy: "Generate a 30-second marketing video for our new brand 'Aurora' anti-aging serum, targeting women over 30, emphasizing 'nighttime repair' and 'brighter skin the next day'."

[0141] User intent parsing: The large model parses the copy into query vectors. This data is then combined with the "30+ women" profile vector and the intent "product efficacy persuasion," and further processed through location encoding to generate a series of storyboard query vectors. (Opening with a question) (Causes of aging) (Product solutions), etc.

[0142] Cross-modal retrieval and aggregation:

[0143] for (Product Solution), the system retrieves information from the dynamic memory network. (Compared to "the moment the product is applied") Text similarity High visual similarity to another "static product display" unit. It is also relatively high.

[0144] After comprehensive scoring, the top 3 results included "rapid absorption and application," "nighttime use scenario," and "next-day comparison effect." Attention weighting. When calculating, the performance-time combined amplification factor is used. They will prefer the recently emerged "fast absorption and application" product category, which has a high conversion rate.

[0145] Finally, the aggregation generates a multimodal context vector. With accompanying causal context (It is emphasized that "effect demonstration" must follow "apply".)

[0146] Causality Preservation Verification: The system checks the generated element sequence [problem display, causes of aging, product application, ?]. If the process ends abruptly after "product application" and the causal preservation score is low, a re-search is triggered, inserting an "effect comparison" element to ensure the causal chain is complete.

[0147] S4.: Causal constraint diffusion model video generation.

[0148] Multi-condition control: Break it down into visual conditions (serum bottle design, dropper, skin texture), textual conditions ("Brighten overnight, witness youth"), and structural conditions (slow-motion close-up + highlight flash).

[0149] Causal consistency constraint: The system uses a pre-trained Stable Video Diffusion model as its foundation. During fine-tuning or inference, the total loss function... .in, The generated video frame sequence is required to ensure that the pixel-level changes from "smearing action" to "skin gloss change" conform to physical laws and causal logic, so as to avoid causal errors such as "becoming brighter without smearing".

[0150] Video assembly: The generated multiple storyboards (total duration 28 seconds) are spliced ​​together in narrative order, and preset background music and transition effects are added to output the final video. .

[0151] S5. Dual-loop reinforcement learning optimization.

[0152] Inner loop optimization (content quality): The denoising process of the diffusion model is treated as an MDP. State The current noisy video latent code, action This represents the noise in the prediction. After video generation, part of it is determined by internal aesthetics and a correlation model. One part is scored (e.g., 0.75), and the other part is submitted for manual evaluation. (e.g., 4.2 / 5). Advantage function The difference between the weighted sum of the two and the baseline is calculated, and the parameters of the diffusion model are updated through policy gradient to make it more inclined to generate high-aesthetic and high-relevance videos.

[0153] Outer Loop Optimization (Publishing Strategy): The publishing strategy network (a PPO Agent) determines the following after video generation: publishing platform (Douyin / Kuaishou / Xiaohongshu), publishing time (20:00 / 12:00), and budget (500 / 1000 RMB). After publishing, the actual CPM (cost per thousand impressions) and ROI are used as immediate rewards. Meanwhile, a RAG-enhanced future value estimator retrieves the long-term performance of similar videos from historical cases, helping to calculate more accurate future value. PPO clipping loss function Used for stable update release strategies.

[0154] S6.RAG enhances autonomous release and system self-evolution.

[0155] RAG Enhanced Publishing Strategy Retrieval: Extraction Feature tags The RAG system retrieved information from its publishing strategy knowledge base: videos posted on Xiaohongshu between 8:00 PM and 10:00 PM, accompanied by hashtags like #NighttimeSkincareRitual and #AntiAgingEssence, had the highest conversion rates. The LLM system then generated the final strategy based on this information.

[0156] Multi-armed slot machine platform selection: The system hesitates between three platforms. Platform A (TikTok) historical average payout. High, but number of explorations There are also many; although platform B (Xiaohongshu) has slightly lower average rewards, its RAG knowledge information gain is higher. High (indicating that, based on historical cases, anti-aging serums have great potential on this platform). The multi-armed slot machine formula is achieved by adding... Ultimately, Xiaohongshu was chosen as the optimal platform. .

[0157] Closed-loop feedback: 24 hours after the video was published on Xiaohongshu, the system collected actual data: click-through rate 12%, completion rate 40%, conversion rate 6%, and share rate 3%. These performance metrics... Feedback was sent back to the system:

[0158] Update memory network: All the scenes in the video are encapsulated in a new... The unit incorporates a dynamic memory network, particularly strengthening the causal relationship between "popularizing anti-aging ingredients and writing about age anxiety."

[0159] Fine-tuning the generation model: Inner-loop reinforcement learning uses high-conversion videos as positive samples to further optimize the generation strategy.

[0160] Updated release strategy library: The successful release strategy has been stored in the RAG knowledge base for future reference for similar products.

[0161] Through the closed-loop iteration of the above 6 steps, the method can continuously learn and evolve from real-time viral content, and autonomously generate and release personalized marketing videos with high causal emergence potential, thereby maximizing the dissemination effect.

[0162] Embodiment 2 of the present invention relates to an AI-based viral video generation and publishing system, the system comprising:

[0163] The causal deconstruction and extraction module is used to decompose the monitored viral videos into atomized genes, which are the smallest reusable creative units, and to establish a causal relationship graph between elements.

[0164] The dynamic memory network module is used to build a real-time evolving knowledge base of trending videos. It encapsulates each atomized gene into a multimodal memory unit and updates the node feature matrix and temporal adjacency matrix dynamically based on new trending video data through a temporal dynamic graph neural network update mechanism.

[0165] The intelligent reorganization module is used to receive user marketing copy, generate dynamic query vectors through large model analysis, perform cross-modal retrieval in the dynamic multimodal memory network, find multiple most similar multimodal memory units, and generate multimodal context vectors through attention-weighted aggregation, while performing causal preservation verification.

[0166] The causal constraint generation module is used to decompose the multimodal context vector into specific generation conditions, input them into the video generator based on the diffusion model, and generate video segments that conform to the causal structure by introducing causal consistency loss and temporal coherence loss.

[0167] The dual-loop reinforcement learning module is used for optimization through a dual-loop reinforcement learning architecture: the inner loop models the video generation process as a Markov decision process and optimizes the generation model parameters using the policy gradient method; the outer loop optimizes the release decision model based on the proximal policy to optimize the release policy.

[0168] The autonomous release and evolution module is used to extract feature tags from the generated video, retrieve the optimal release parameters in the release strategy knowledge base by retrieving enhanced RAGs, and execute autonomous release; collect actual business data after release, and perform closed-loop updates on the temporal dynamic multimodal memory network, diffusion model, and release strategy model.

[0169] Embodiment 3 of the present invention relates to a computer-readable storage medium for storing a computer program, which is executed by a processor to perform the method of Embodiment 1.

[0170] Embodiment 4 of the present invention relates to a computer program product, the computer program product comprising a computer program executed by a processor for performing the method of Embodiment 1.

[0171] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An AI-based blockbuster video generation and publishing method, characterized in that, The method includes the following steps: S1. Decompose the monitored viral videos into atomized genes, which are the smallest reusable creative units, and establish a causal relationship graph between elements; S2. Construct a real-time evolving knowledge base for trending videos, encapsulate each atomized gene as a multimodal memory unit, and dynamically update the node feature matrix and temporal adjacency matrix based on new trending video data through a temporal dynamic graph neural network update mechanism; S3. Receive user marketing copy, generate dynamic query vectors through large model analysis, perform cross-modal retrieval in the dynamic multimodal memory network, find multiple most similar multimodal memory units, generate multimodal context vectors through attention-weighted aggregation, and perform causal preservation verification at the same time; S4. Decompose the multimodal context vector into specific generation conditions, input them into a video generator based on a diffusion model, and generate video segments that conform to the causal structure by introducing causal consistency loss and temporal coherence loss. S5. Optimization is achieved through a dual-loop reinforcement learning architecture: the inner loop models the video generation process as a Markov decision process and optimizes the generation model parameters using the policy gradient method; the outer loop optimizes the release decision model based on the proximal policy to optimize the release policy. S6. Extract feature tags from the generated video, retrieve the optimal release parameters in the release strategy knowledge base by using the enhanced RAG retrieval method, and perform autonomous release; collect actual business data after release, and perform closed-loop updates on the time-series dynamic multimodal memory network, diffusion model, and release strategy model. 2.The AI-based bestseller video generation and publishing method of claim 1, wherein, In step S1, the decomposition specifically includes: using a scene detection algorithm to analyze viral videos. Divided into K storyboard-level units Extract visual feature vectors for each scene. Text feature vectors Motion feature vectors And associate performance tags ; Construct a structural causal model ,in It is a set of atomized gene nodes. For causal relationship, The causal strength weight matrix is ​​used, and the causal emergence effect score for each atomized gene combination is calculated. .

3. The AI-based method for generating and publishing viral videos according to claim 2, characterized in that, In step S1, the causal emergent effect score is calculated The calculation is: ； wherein, is the causal influence strength from element to , is the synergy effect function, is the emergence correction term. 4.The AI-based hit video generation and publishing method of claim 1, wherein, In step S2, the temporal dynamic graph neural network update mechanism is as follows: ； ； in, The element feature matrix at time t+1 is the visual feature vector + text feature vector + motion feature vector. Let be the element characteristic matrix at time t; Let be the temporal adjacency matrix at time t, representing the dynamic correlation strength between popular elements; Let be the node attribute matrix at time t, containing the performance labels of the elements; These are learnable parameters; For graph neural networks; The temporal residual weight is used to control the proportion of historical memory retention. The map update rate controls the speed at which new knowledge is absorbed; Let be the temporal adjacency matrix at time t+1; Let be the query matrix at time t; The key matrix at time t+1; This is the value matrix at time t+1; For attention head dimension. 5.The AI-based best-seller video generation and publishing method of claim 1, wherein, In step S3, generating the dynamic query vector specifically includes the following calculations: ； Parameter Q is the base query vector; Embedding functions for large language models; Marketing copy text entered by the user; User profile vectors; This is the intent recognition vector; This is a vector concatenation function; ； parameter The position-aware query vector for the j-th scene; Based on the query projection matrix; This is a position-encoded vector; The position is encoded as a projection matrix; The set of query vectors mentioned above represents the set of generated storyboard query vectors; This is the encoding function for the preceding context; This is the context projection matrix. 6.The AI-based hit video generation and publishing method of claim 1, wherein, In step S2, cross-modal retrieval is performed in the dynamic multimodal memory network to find multiple most similar multimodal memory units. The cross-modal retrieval similarity scoring formula is as follows: ； in, For query vector With memory unit The overall similarity score; The cosine similarity function; For the r-th candidate memory unit, it is one of the Top-k candidates retrieved from the dynamic memory network; The text projection matrix of the position-aware query vector of the j-th scene; The visual projection matrix of the position-aware query vector for the j-th scene; The motion projection matrix for the position-aware query vector of the j-th scene; Let be the text feature vector of the r-th candidate memory unit; Let be the visual feature vector of the r-th candidate memory unit; Let r be the motion feature vector of the r-th candidate memory unit; These are the modal weighting coefficients. 7.The AI-based bestseller video generation and publishing method of claim 1, wherein, In step S5, the formula for calculating the policy gradient of the inner loop optimization is: ； in, This represents the gradient of the policy objective function; Let the objective function be the policy objective function; To generate model parameters; This refers to the state-action sequence trajectory during the video generation process. To generate a strategy, determine the action to take in a given state; Calculate the expectation of all possible trajectories according to the policy distribution; For time steps; The state at time t; The action taken at time t; Estimate the content advantage function; The final generated video; These are the weighting coefficients for the prediction model; Scoring of the automatic prediction model; Baseline value; The score is determined by human evaluation; the content advantage function is... Score assessed by humans With automatic prediction model scoring The weighted combination is obtained by subtracting the baseline value.

8. An AI-based viral video generation and publishing system, wherein the system uses the AI-based viral video generation and publishing method according to any one of claims 1-7, characterized in that, The system includes: The causal deconstruction and extraction module is used to decompose the monitored viral videos into atomized genes, which are the smallest reusable creative units, and to establish a causal relationship graph between elements. The dynamic memory network module is used to build a real-time evolving knowledge base of trending videos. It encapsulates each atomized gene into a multimodal memory unit and updates the node feature matrix and temporal adjacency matrix dynamically based on new trending video data through a temporal dynamic graph neural network update mechanism. The intelligent reorganization module is used to receive user marketing copy, generate dynamic query vectors through large model analysis, perform cross-modal retrieval in the dynamic multimodal memory network, find multiple most similar multimodal memory units, and generate multimodal context vectors through attention-weighted aggregation, while performing causal preservation verification. The causal constraint generation module is used to decompose the multimodal context vector into specific generation conditions, input them into the video generator based on the diffusion model, and generate video segments that conform to the causal structure by introducing causal consistency loss and temporal coherence loss. The dual-loop reinforcement learning module is used for optimization through a dual-loop reinforcement learning architecture: the inner loop models the video generation process as a Markov decision process and optimizes the generation model parameters using the policy gradient method; the outer loop optimizes the release decision model based on the proximal policy to optimize the release policy. The autonomous release and evolution module is used to extract feature tags from the generated video, retrieve the optimal release parameters in the release strategy knowledge base by retrieving enhanced RAGs, and execute autonomous release; collect actual business data after release, and perform closed-loop updates on the temporal dynamic multimodal memory network, diffusion model, and release strategy model.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program, which is executed by a processor to perform the AI-based method for generating and publishing viral videos according to any one of claims 1-7.

10. A computer program product, characterised in that, The computer program product includes a computer program that is executed by a processor to perform the AI-based method for generating and publishing viral videos as described in any one of claims 1-7.

Images