A video propaganda product intelligent design method based on a large model and a knowledge base

Abstract

The application belongs to the technical field of large model application, and particularly relates to a video propaganda product intelligent design method based on a large model and a knowledge base. In view of the problems of long design cycle, low standardization of traditional video propaganda product design, and frame motion confusion, content illusion, weak demand analysis and the like of the existing large model video generation, a video propaganda product intelligent design method based on a large model and a knowledge base is provided. The method first preprocesses the user natural language demand to generate a standardized structured data set; a video script is generated by a script generation large model, and is segmented into a split screen script with time sequence anchor points; an interframe motion optical flow field is generated by an optical flow prediction model to constrain the video large model to generate coherent split screen frames; after video splicing, abnormality checking is carried out, and for abnormal segments, the split screen script is regenerated and replaced through reverse matching, until the output product passes the checking. The application improves the video production efficiency and demand matching degree, and guarantees smooth picture and real and compliant content.

Description

Technical Field

[0001] This invention belongs to the field of large model application technology, and in particular relates to an intelligent design method for video promotional materials based on large models and knowledge bases. Background Technology

[0002] With the deep integration of artificial intelligence big data models and digital media technologies, video promotional materials have become a core form of communication for corporate brand promotion and product marketing. Traditional video promotional material design requires manual completion of the entire process, including scriptwriting, storyboarding, video production, and post-production verification. This results in problems such as long production cycles, high labor costs, and low standardization of creative concepts, making it difficult to meet the current demand for efficient, high-volume promotional material production.

[0003] While current video generation technology based on large models has been attempted to be applied to the field of promotional material design, it suffers from significant technical shortcomings. Firstly, the lack of inter-frame motion constraints results in poor continuity of main subject movements, chaotic inter-frame temporal logic, and an overall lack of smoothness that fails to meet the standards for promotional material playback. Secondly, large models are prone to generating content illusions, with generated content violating physical rules and promotional facts, making it difficult to guarantee the authenticity and compliance of information. Thirdly, the generation process has extremely low controllability, making it impossible to precisely adjust the storyboard, timing, and visual elements according to user needs, resulting in significant difficulties in accurate adaptation. Fourthly, its ability to interpret natural language requirements is weak, failing to translate vague user needs into standardized design instructions, leading to a large discrepancy between the final product and actual requirements. Existing technologies cannot simultaneously achieve efficiency, continuity, and accuracy in video generation, severely hindering the large-scale implementation of intelligent design for video promotional materials. Therefore, there is an urgent need to develop an efficient, accurate, and controllable intelligent design method. Summary of the Invention

[0004] To address the technical problems existing in the above-mentioned background art, this invention proposes an intelligent design method for video promotional materials based on a large model and knowledge base.

[0005] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0006] S1. Receive user input for video promotional material design natural language requirements, perform preprocessing operations to decompose the natural language requirements into a standardized structured dataset of promotional requirements;

[0007] S2. Input the structured dataset into the script to generate a large model and generate a video script;

[0008] S3. Perform a segmentation operation on the video script, dividing it into multiple storyboard scripts with time-series anchor points;

[0009] S4. Based on each storyboard script, generate consecutive video frames of each storyboard in the order of the time anchor points; when generating a single storyboard, use the main action trajectory of the storyboard as a hard temporal constraint, generate an inter-frame motion optical flow field through an optical flow prediction model, and constrain the large video model to generate consecutive frames of the storyboard.

[0010] S5. Concatenate the video frame sequences of each storyboard script in chronological order to obtain the initial video sequence, perform anomaly verification on the initial video sequence, and extract abnormal segments.

[0011] S6. For the time sequence anchor point to which the abnormal segment belongs, reverse match the corresponding original storyboard script, regenerate the storyboard script to replace the original storyboard script, until the initial video passes all checks, and output the finished video promotional product that meets the requirements.

[0012] Preferably, the specific implementation of the preprocessing operation to decompose natural language requirements into a standardized structured dataset of publicity requirements includes:

[0013] S11. Perform a full character scan on the natural language requirement text for video promotional material design input by the user, and remove invisible characters, special symbols, invalid spaces, and duplicate or redundant characters to obtain the initial screening requirement text.

[0014] S12. Perform semantic filtering on the initial screened text to remove irrelevant modifiers and obtain the semantically filtered text.

[0015] S13. Finally, intent recognition is performed, and the semantically filtered text is compared with standardized professional text in the preset industry knowledge base. The matching degree is calculated and a standardized structured dataset is generated.

[0016] Preferably, the specific implementation of semantic filtering of the initially screened text and removal of irrelevant modifiers to obtain the semantically filtered text in step S12 is as follows:

[0017] S121. Input the initial screened text into the semantic field construction module of the large model, generate a global semantic field with whole sentences as the carrier, map all words or phrases in the text to semantic nodes in the semantic field, and form a continuous semantic field distribution.

[0018] S122. Through the gradient calculation of the semantic field of the large model, the gradient value in the semantic field is automatically identified. Nodes with gradient values ​​higher than the set gradient threshold are marked as core semantic nodes, and the remaining nodes are modifier nodes.

[0019] S123. Calculate the semantic field association gradient K between the modifier node and the core semantic node. The semantic field association gradient K represents the absolute gradient value of the mean gradient of the core semantic node and the mean gradient of the modifier node, and adaptively generate a gradient filtering threshold. ,in, These are the upper and lower bounds of the semantic field gradient, respectively. We traverse all modifier nodes, remove irrelevant modifiers whose associated gradients are less than the adaptive gradient filtering threshold, and output the semantically filtered text without redundancy.

[0020] Preferably, in step S13, intent recognition is performed, and the semantically filtered text is compared with standardized professional text in a preset industry knowledge base to calculate the matching degree and finally generate a standardized structured dataset. The specific implementation of this is as follows:

[0021] S131. Perform semantic segmentation on the semantically filtered text, breaking it down into several minimum intent semantic units;

[0022] S132. Input several minimum intent semantic units into the domain adaptive pre-trained BERT model to generate intent semantic unit feature vectors;

[0023] S133. Construct a standardized semantic text library. Perform cosine similarity matching between the intention semantic unit feature vector and the full set of standardized semantic feature vectors in the standardized professional text library. Calculate the matching score and map all semantically filtered texts to the standardized semantics with the highest matching score to obtain a standardized structured dataset.

[0024] Preferably, step S2 involves inputting the structured dataset into a script to generate a large model. The specific implementation of generating the video script includes:

[0025] S21. Construct a knowledge base specific to the promotional materials domain, and pre-encode all knowledge entries in the knowledge base to generate corresponding semantic feature vectors, denoted as . ,in, Let be the feature vector corresponding to the j-th knowledge entry in the knowledge base, and m be the total number of entries in the knowledge base;

[0026] S22. Denote the obtained standardized publicity demand structured dataset as D, where D contains n dimensions of demand feature fields. The required global feature vector is generated by encoding the global features of D using a domain-pretrained BERT encoding model. Calculate the global feature vector required Feature vectors of each entry in the knowledge base The cosine similarity is used to filter out knowledge items with matching scores higher than the preset value to generate a prior constraint feature set;

[0027] S23. Generate prompt words based on structured datasets and prior constraint feature sets;

[0028] S24. Input the prompt word "Prompt" into the pre-fine-tuned script to generate a large model and generate an initial candidate script set. The generation stage uses a beam search algorithm, sets the beam width to N, and generates N initial candidate scripts to form a candidate script set. For each candidate script, a conditional probability is calculated, and the candidate script with the highest conditional probability is selected as the baseline script; a standardized format encapsulation is performed to generate a video script; the conditional probability is the candidate script. The model generation probability under the prompt word Prompt constraint.

[0029] Preferably, step S3, which involves performing a segmentation operation on the video script, is implemented as follows:

[0030] S31. Perform global feature extraction on the video script, extracting the semantic features and scene features of each set time period in the video script;

[0031] S32. Calculate the correlation between semantic features and scene features within two consecutive set time periods, and obtain the normalized correlation value by weighted summation;

[0032] S33. Preset correlation threshold: When the correlation value of two consecutive time periods is lower than the correlation threshold, it is automatically determined as the corresponding segmentation benchmark point to perform the segmentation operation, and the initial time of each storyboard is marked as the time sequence anchor point.

[0033] Preferably, in step S4, when generating a single scene, the main action trajectory of the scene is used as a temporal hard constraint. An inter-frame motion optical flow field is generated through an optical flow prediction model to constrain the generation of continuous frames of the scene by the large video model. The specific implementation of this constraint is as follows:

[0034] S41. Decompose the main body motion trajectory in the storyboard into a continuous set of time nodes according to the time anchor points, bind the main body posture coordinates and motion velocity to each time node, and construct the motion trajectory constraint matrix M.

[0035] S42. Input the core elements of the scene and the motion trajectory temporal constraint matrix M from the storyboard into the optical flow prediction model. Initialize the inter-frame initial optical flow field guided by the trajectory gradient, and generate the target motion optical flow field through constraint iteration optimization. ,in, Let be the initial optical flow field tensor. The coordinates of any pixel point in the plane of each frame image. The trajectory guidance weights have values ​​ranging from 0 to 1. To calculate the spatial gradient of matrix M, the subject motion trend information in the motion trajectory temporal constraint matrix M is transformed into a correction value for the optical flow field.

[0036] S43. Decompose the target motion optical flow field into frame-by-frame motion constraint vector groups, and convert the motion constraint vector groups into inter-frame motion hard constraint instructions that can be parsed by the large video model, thereby locking the motion direction and motion speed of each frame pixel.

[0037] S44. The motion hard constraint command is generated as an immutable forced condition between frames, driving the model to strictly follow the pixel motion trajectory defined by the optical flow field to generate continuous video frames of the storyboard frame by frame.

[0038] Preferably, in step S5, the abnormal segment is extracted by performing anomaly verification on the initial video sequence and breaking down the initial video sequence into continuous image frames, performing factual logic verification on the continuous image frames, identifying content that violates physical rules and the standard of propaganda facts, and determining it as an abnormal segment.

[0039] Compared with existing technologies, the advantages and positive effects of this invention are as follows: This invention achieves intelligent design of video promotional materials throughout the entire process, solving the problems of long production cycles, high costs, and low standardization in traditional production, and significantly improving production efficiency. Through structured parsing of natural language requirements, it accurately transforms ambiguous requirements into standard instructions, reducing the deviation between the finished product and the requirements. Based on an optical flow prediction model, it constructs hard constraints for inter-frame motion, solving the problems of chaotic subject movement trajectories and insufficient image smoothness. Combining a domain knowledge base and anomaly verification mechanism, it avoids content illusions caused by large models, ensuring the compliance and authenticity of promotional content. Simultaneously, scene segmentation and temporal anchoring improve the controllability of generation, enabling efficient batch production and promoting the large-scale implementation of intelligent design for video promotional materials. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart illustrating an intelligent design method for video promotional materials based on a large model and knowledge base. Detailed Implementation

[0042] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0043] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0044] In this embodiment, to address issues such as chaotic inter-frame temporal logic during video generation, the tendency of large models to generate content illusions, and extremely low controllability of the generation process, this invention provides an intelligent design method for video promotional materials based on large models and knowledge bases. The specific implementation is as follows: Figure 1 As shown.

[0045] S1. Receive user input for video promotional material design natural language requirements, perform preprocessing operations to decompose the natural language requirements into a standardized structured dataset of promotional requirements.

[0046] Furthermore, the specific implementation of the preprocessing operation to decompose natural language requirements into a standardized structured dataset of promotional requirements includes:

[0047] The system performs a full character scan on the user-input natural language requirement text for video promotional materials, removing invisible characters, special symbols, invalid spaces, and redundant characters to obtain the initial requirement text. Specifically, a full character scanning engine is activated, traversing every character unit of the requirement text, establishing a character index mapping table, and accurately locating invisible characters (such as newlines, tabs, and null characters), special symbols (such as emoticons, garbled characters, and punctuation marks unrelated to the promotion), invalid spaces (such as consecutive spaces, leading and trailing spaces, and extra spaces between paragraphs), and redundant characters (such as repeated words, repeated sentences, and meaningless reduplicated words). During the scanning process, a combination of regular expression matching and character encoding verification is used, performing dual checks on ASCII and Unicode codes to ensure no abnormal characters are missed. After recognition, a batch removal operation is performed, retaining valid Chinese characters, numbers, standard punctuation marks, and core promotional vocabulary. Simultaneously, the text format is standardized, the character encoding is unified to UTF-8, and the text paragraph structure is adjusted to single-line continuous text. After removal and standardization, the initial requirement text is output.

[0048] The initial screening text undergoes semantic filtering to remove irrelevant modifiers, resulting in semantically filtered text. Further, the initial screening text is input into the semantic field construction module of a large model. A global semantic field is generated using complete sentences as the basis, mapping all words or phrases within the text to semantic nodes in the semantic field, forming a continuous semantic field distribution. Specifically, the initial screening text undergoes sentence-by-sentence preprocessing, using natural sentences as the smallest processing unit. Accurate sentence segmentation is achieved through Chinese punctuation mark boundary recognition, ensuring that each processing unit possesses complete and independent semantic logic. Subsequently, the segmented complete sentence text is input into a BERT encoding model finely tuned from a corpus of millions of promotional material designs. Word embedding is performed through the model's 12-layer Transformer encoder, generating a 768-dimensional semantic feature vector for each word and phrase in the text, while simultaneously capturing the contextual relationships and semantic dependencies between words. Based on the semantic feature vectors, a semantic field space with the complete sentence semantics as the global boundary is constructed. Each word or phrase is mapped to an independent semantic node in the semantic field. The node coordinates are determined by the principal component dimensionality reduction result of the semantic feature vectors, and the node weights are assigned by the semantic contribution of the word in the sentence. Finally, by fitting the spatial distribution of semantic nodes with their associated weights, a continuous and complete global semantic field distribution is formed.

[0049] Through large-scale semantic field gradient calculation, gradient values ​​in the semantic field are automatically identified. Nodes with gradient values ​​higher than a set gradient threshold are designated as core semantic nodes, while the remaining nodes are designated as modifier nodes. Specifically, based on the constructed global semantic field, the semantic field gradient value is defined with the information gain of semantic nodes on the core intent of the promotional needs as the core. The gradient value is used to characterize the coreness and information weight of semantic nodes in the global semantic field. During gradient calculation, the gradient value is weighted based on the cosine similarity between the feature vector of the semantic node and the global semantic feature vector of the entire sentence, combined with the weight coefficients of the preset core keyword vocabulary in the promotional domain, to ensure a strong correlation between the gradient value and the promotional needs. Subsequently, based on the gradient value distribution of all nodes in the semantic field, a 3D model is used to calculate the gradient value. The criteria set gradient thresholds to ensure that the thresholds are statistically reasonable and applicable to various scenarios. All semantic nodes are traversed, and nodes with gradient values ​​higher than the set gradient threshold are marked as core semantic nodes. These nodes correspond to core needs information such as the advertised product, key selling points, and target audience. The remaining nodes are marked as modifier nodes, thus completing the precise division between the two types of nodes.

[0050] Calculate the semantic field association gradient K between the modifier node and the core semantic node. The semantic field association gradient K represents the absolute gradient value of the mean gradient of the core semantic node and the mean gradient of the modifier node. Adaptively generate a gradient filtering threshold. ,in, These are the upper and lower bounds of the semantic field gradient, respectively. We traverse all modifier nodes, remove irrelevant modifiers whose associated gradients are less than the adaptive gradient filtering threshold, and output the semantically filtered text without redundancy.

[0051] Finally, intent recognition is performed, comparing the semantically filtered text with standardized professional text in a pre-set industry knowledge base, calculating the matching degree, and ultimately generating a standardized structured dataset. Further, the semantically filtered text is semantically segmented into several smallest semantic intent units. Specifically, lexical and syntactic preprocessing is performed on the semantically filtered text. A word segmentation model loaded with a professional thesaurus for promotional materials is used for accurate word segmentation, simultaneously performing part-of-speech tagging and dependency parsing to identify the core subject-verb-object structure and modifier relationships, clarifying the semantic logical hierarchy. Then, based on a pre-trained BIO sequence labeling model for the promotional domain, semantic boundary recognition is performed on the preprocessed text, accurately labeling the start, internal, and end positions of each independent intent unit, completing the decomposition.

[0052] Next, several minimal intent semantic units are input into a domain-adaptive pre-trained BERT model to generate intent semantic unit feature vectors. Specifically, the domain-adaptive pre-trained and optimized BERT model performs deep semantic encoding on the decomposed minimal intent semantic units, generating highly discriminative and representative intent semantic unit feature vectors, realizing the vectorized and computable transformation of semantic information. First, the domain-adaptive adaptation of the base model is completed. Using the Chinese BERT-base model as a foundation, further pre-training is performed based on the knowledge base, optimizing the masked language model and sentence-level semantic matching tasks, allowing the model to fully learn the semantic features, terminology associations, and scene logic of the promotional material design domain, strengthening its ability to represent the intention of promotional needs. Subsequently, standardization preprocessing is performed on each input minimal intent semantic unit, adding [CLS] and [SEP] special markers to complete tokenization conversion, generating a model-recognizable token sequence, segment sequence, and attention mask sequence, unifying the sequence length to 128 dimensions to adapt to the model input specifications. Finally, the preprocessed sequence is input into the adapted BERT model. After multi-head attention calculation and feedforward network processing by a 12-layer Transformer encoder, the output of the last hidden layer corresponding to the [CLS] label is extracted to generate a 768-dimensional intention semantic unit feature vector.

[0053] Finally, a standardized semantic text library is constructed. The feature vectors of intent semantic units are matched with the full set of standardized semantic feature vectors in the standardized professional text library using cosine similarity. Matching scores are calculated, and the semantically filtered texts are mapped to the standardized semantics with the highest matching score, resulting in a standardized structured dataset. Specifically, a domain-adaptive BERT model, identical to the steps described above, is used to pre-encode each standardized semantic text in the library, generating corresponding standardized semantic feature vectors, which are stored in a vector database and a fast retrieval index is established. Then, the feature vector of each intent semantic unit is matched with the full set of standardized semantic feature vectors of the corresponding dimension in the vector database using cosine similarity calculation. The matching score ranges from 0 to 1, with higher scores indicating stronger semantic fit. For each intent unit, the standardized semantic with the highest matching score is selected as the final mapping result, completing the standardization transformation of all intent units. Finally, according to a pre-defined structured dataset template, standardized semantics for each dimension are filled into the corresponding fields, and field constraints and parameter attributes are added to generate a standardized structured dataset of promotional needs with complete fields and a unified format.

[0054] S2. Input the structured dataset into the script to generate a large model and then generate the video script. Specifically, construct a knowledge base specific to the promotional materials domain, pre-encode all knowledge entries in the knowledge base to generate corresponding semantic feature vectors, denoted as... ,in, Let m be the feature vector corresponding to the j-th knowledge entry in the knowledge base, and m be the total number of entries in the knowledge base; let D be the structured dataset of standardized publicity needs obtained, which contains n-dimensional demand feature fields. The required global feature vector is generated by encoding the global features of D using a domain-pretrained BERT encoding model. Calculate the global feature vector required Feature vectors of each entry in the knowledge base The cosine similarity is used to filter out knowledge items with matching scores higher than the preset value to generate a prior constraint feature set;

[0055] Based on a structured dataset and a prior constraint feature set, a prompt word is generated. A fixed hierarchical architecture for the prompt is built, consisting of four modules: a core requirement layer, a domain constraint layer, a format specification layer, and an output requirement layer. The n-dimensional feature fields of the structured dataset D are written one by one to fully recreate the user's comprehensive promotional needs. The domain constraint layer module integrates highly relevant knowledge items from the prior constraint feature set according to three categories: compliance red lines, script paradigms, and scenario adaptation, extracting core constraint rules and writing them into the module to clearly define the boundaries of prohibited content. The format specification layer module defines the standardized format of the video script, requiring it to include seven fixed fields: script name, total promotional duration, core theme, scene number, visual content, dialogue / narration, and single-segment duration annotation. The output requirement layer module defines the script's language style, word count limits, and logical coherence requirements. Finally, the content of the four modules is integrated to generate a complete prompt word that can be directly input into a large model.

[0056] The prompt word "Prompt" is input into a pre-fine-tuned script to generate a large model and an initial candidate script set. The generation stage employs a beam search algorithm, setting the beam width to N, to generate N initial candidate scripts that form the candidate script set. For each candidate script, a conditional probability is calculated, and the candidate script with the highest conditional probability is selected as the baseline script; a standardized format encapsulation is performed to generate a video script; the conditional probability is the candidate script. The model generates probabilities under the Prompt constraint. Specifically, a large model is generated using a script. The completed Prompt is input into a pre-tuned script to generate the large model. During the generation stage, a beam search algorithm is used to optimize the generation effect. The beam width N is set to 5, meaning that at each step of token generation, the 5 candidate sequences with the highest probability are retained. Five parallel generation paths are maintained throughout the process, ultimately generating 5 complete initial candidate scripts, forming a candidate script set. Then, the conditional probability is calculated for each candidate script. This probability is the joint generation probability of the entire sequence of tokens in the candidate script under the Prompt constraint. The higher the probability value, the stronger the fit between the script and the constraint requirements. All candidate scripts are traversed, and the script with the highest conditional probability is selected as the baseline script. Finally, the baseline script is standardized and encapsulated according to industry standards, including format validation, field completion, and timing regularization, to generate a video script that is formatted correctly, compliant with content, and fully matches the user's needs.

[0057] S3. Perform a segmentation operation on the video script, dividing it into multiple storyboard scripts with temporal anchor points. Specifically, the segmentation operation involves extracting global features from the video script, extracting semantic and scene features for each set time period; calculating the correlation between semantic and scene features within two consecutive set time periods, and obtaining a normalized correlation value through weighted summation; setting a correlation threshold, and automatically determining the corresponding segmentation benchmark point when the correlation value between two consecutive set time periods is lower than the threshold, and marking the initial time of each storyboard script as a temporal anchor point.

[0058] Specifically, the generated standardized video script is first subjected to full structured parsing to accurately extract all information, including total duration, segment duration annotations, scene descriptions, visual content, dialogue narration, and camera movement requirements, thus binding and mapping the script text to temporal information. Then, using 1 second as the smallest unit, the entire duration of the video script is divided into consecutive, equally long time segments. Each time segment is bound to script text content within its corresponding temporal range, achieving temporal slicing of the entire script. For the bound text of each time segment, a two-dimensional extraction of semantic and scene features is performed: semantic feature extraction uses a domain-adaptive pre-trained BERT model, which is consistent with the requirement encoding, to generate a 768-dimensional semantic feature vector, capturing the semantic connotations of narrative logic and core promotional information within the time segment; scene feature extraction uses a pre-trained scene entity sequence annotation model to extract core elements such as scene location, main objects, shot size, camera movement, and visual style, encoding them to generate a scene feature vector of the same dimension. Finally, full-time, full-coverage feature extraction is completed, providing standardized input for subsequent correlation calculations. For all consecutive time-series pairs, semantic and scene feature vectors from the preceding time period t and the following time period t+1 are retrieved for dual-dimensional correlation calculation. Semantic correlation calculation uses a cosine similarity algorithm to calculate the cosine value of the semantic feature vectors of the two time periods, obtaining a semantic correlation score ranging from 0 to 1. A higher score indicates stronger narrative logic and coherence of promotional information between the two time periods. Similarly, scene correlation calculation uses cosine similarity to calculate the matching degree of scene feature vectors between the two time periods, obtaining a scene correlation score. A higher score indicates stronger consistency of visual scenes between the two time periods. Then, weighted weights are set based on the creation patterns of promotional videos: a weight coefficient of 0.6 for scene correlation and a weight coefficient of 0.4 for semantic correlation, with the sum of the two weights being 1. The original correlation values ​​are obtained through weighted summation. Finally, min-max normalization is performed on all original correlation values, mapping all values ​​to the 0-1 range, eliminating numerical distribution bias, and generating a standardized full-time correlation sequence. Finally, based on the statistical distribution of the normalized correlation sequence throughout the entire time series, the lower quartile Q1 of the sequence is calculated as an adaptive baseline threshold. Simultaneously, a fixed minimum threshold of 0.3, conforming to the industry standard for promotional video storyboards, is set. The maximum of these two values ​​is taken as the final segmentation threshold. Subsequently, all adjacent time period pairs are traversed throughout the entire time series, and their corresponding correlation values ​​are compared with the threshold. When the correlation value of an adjacent time period pair is lower than the threshold, the time boundary between the two time periods is automatically marked as the storyboard segmentation baseline point, i.e., the end node of the previous storyboard and the start node of the next storyboard. Based on all segmentation baseline points, the complete video script is split into multiple independent storyboard scripts. Each storyboard script corresponds to a continuous and uninterrupted time interval, and the initial time of each storyboard script is marked as a unique time anchor point. This process ultimately completes the storyboard segmentation and time anchoring that conforms to industry standards.

[0059] S4. Based on each storyboard script, generate consecutive video frames of each storyboard in the order of the time anchor points; when generating a single storyboard, use the main action trajectory of the storyboard as a hard temporal constraint, generate the inter-frame motion optical flow field through the optical flow prediction model, and constrain the large video model to generate consecutive frames of the storyboard.

[0060] Specifically, the subject's motion trajectory within the storyboard is decomposed into a continuous set of temporal nodes according to temporal anchor points. The subject's posture coordinates and motion velocity are bound to each temporal node to construct the motion trajectory constraint matrix M.

[0061] The core elements of the scene in the storyboard and the temporal constraint matrix M of the motion trajectory are input into the optical flow prediction model. The initial optical flow field between frames is initialized with the trajectory gradient as a guide, and the target motion optical flow field is generated through constraint iteration optimization. ,in, Let be the initial optical flow field tensor. The coordinates of any pixel point in the plane of each frame image. The trajectory guidance weights have values ​​ranging from 0 to 1. To calculate the spatial gradient of matrix M, the subject's motion trend information in the temporal constraint matrix M of the motion trajectory is transformed into a correction quantity for the optical flow field. It should be noted that this invention does not require consideration of dimensions, because... The optical flow field is obtained by preprocessing all pixel vectors by dividing the global maximum optical flow amplitude. The dimensionless normalized optical flow field is obtained. In addition, the matrix M is also normalized to the interval of 0 to 1 according to its own global maximum value. Therefore, the formula for generating the target motion optical flow field does not have a dimension problem. After generating the final optical flow field, it is reverse normalized according to the original maximum amplitude value to restore the actual usable optical flow field and drive the generation of the large video model.

[0062] The target motion optical flow field is decomposed into frame-by-frame motion constraint vector groups. These motion constraint vector groups are then converted into hard inter-frame motion constraints that can be parsed by the large video model, locking the motion direction and velocity of pixels in each frame. Specifically, the target motion optical flow field is decomposed temporally. The target motion optical flow field is a four-dimensional tensor, including the total number of frames in the sequence, the height and width pixel values ​​of the generated image, and the last two dimensions corresponding to the motion vectors in the x and y directions of the pixels, respectively. Following the temporal sequence, the optical flow field is decomposed into frame-by-frame motion constraint vector groups. Each vector group corresponds to the full-screen pixel motion parameters from frame t to frame t+1, completely covering the motion direction and velocity of all pixels in the image. Subsequently, based on the subject's motion trajectory constraint matrix, the vector groups are weighted regionally. The motion vectors of the core subject region are designated as unmodifiable mandatory constraints, while the background region vectors are designated as reference constraints. Finally, the frame-by-frame motion constraint vector groups are converted into hard constraint commands that can be parsed by the large video model.

[0063] Motion hard constraints are applied as immutable, mandatory conditions for inter-frame generation, driving the model to strictly adhere to pixel motion trajectories defined by the optical flow field, generating consecutive video frames of the storyboard frame by frame. Specifically, the generation rules of the large-scale text-based video model, pre-fine-tuned in the promotional video domain, are configured at the lowest level. The generation priority of motion hard constraints is set to the highest level, with a fixed constraint weight coefficient of 1.0. This priority is higher than all soft constraint inputs such as text prompts, visual style, and color tone, prohibiting any pixel motion generation that deviates from the constraint instructions from the model's generation logic level. Subsequently, the autoregressive generation process of the storyboard frame by frame is initiated. Using the baseline image of the first frame of the storyboard as the initial input, the first set of inter-frame motion hard constraints is simultaneously input into the model's ControlNet control module and semantic encoding layer, generating the second frame through dual constraints. After each frame is generated, it is used as the initial input for the next frame, matching the corresponding temporal hard constraint instructions, and so on, to complete the generation of all storyboard frames. A real-time deviation verification mechanism is embedded in the generation process. For each frame generated, the actual pixel motion vector of the frame and the previous frame are calculated and compared with the constraint vector group. Frames with deviations exceeding the preset threshold are immediately regenerated, and the final output is a continuous video frame of the storyboard that fully conforms to the motion constraints.

[0064] S5. Concatenate the video frame sequences of each storyboard in chronological order to obtain the initial video sequence. Perform anomaly checks on the initial video sequence and extract abnormal segments. Specifically, break down the initial video sequence into consecutive image frames, perform factual logic checks on the consecutive image frames, identify content that violates physical rules and propaganda standards, and determine it as an abnormal segment.

[0065] S6. For the time sequence anchor point to which the abnormal segment belongs, reverse match the corresponding original storyboard script, regenerate the storyboard script to replace the original storyboard script, until the initial video passes all checks, and output the finished video promotional product that meets the requirements.

[0066] Specifically, the temporal anchor points corresponding to each anomalous segment are extracted, and the original storyboard corresponding to the anomalous segment is accurately back-matched. Then, the reasons for the anomalies in the original storyboard are integrated into regeneration prompts. These prompts are input into a large-scale model to specifically correct the anomalies, generating an optimized storyboard that replaces the original version. Following the frame generation process, a new sequence of storyboard video frames conforming to motion constraints is generated based on the new storyboard, replacing the anomalous segments at the corresponding temporal anchor points in the initial video. Finally, a full anomaly check is performed on the repaired video. If anomalies still exist, the process is repeated until the video passes all checks. Ultimately, format standardization and final product packaging are completed, resulting in a compliant and compliant video promotional material.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent design of video promotional materials based on a large model and knowledge base, characterized in that, Includes the following steps: S1. Receive user input for video promotional material design natural language requirements, perform preprocessing operations to decompose the natural language requirements into a standardized structured dataset of promotional requirements; S2. Input the structured dataset into the script to generate a large model and generate a video script; S3. Perform a segmentation operation on the video script, dividing it into multiple storyboard scripts with time-series anchor points; S4. Based on each storyboard script, generate consecutive video frames of each storyboard in the order of the time anchor points; when generating a single storyboard, use the main action trajectory of the storyboard as a hard temporal constraint, generate an inter-frame motion optical flow field through an optical flow prediction model, and constrain the large video model to generate consecutive frames of the storyboard. S5. Concatenate the video frame sequences of each storyboard in chronological order to obtain the initial video sequence, perform anomaly checks on the initial video sequence, and extract abnormal segments. S6. For the time anchor point to which the abnormal segment belongs, reverse match the corresponding original storyboard script, regenerate the storyboard script to replace the original storyboard script, until the initial video passes all checks and outputs the finished video promotional product that meets the requirements. Step S2 involves inputting the structured dataset into a script to generate a large model. The specific implementation of generating the video script includes: S21. Construct a knowledge base specific to the promotional materials domain, and pre-encode all knowledge entries in the knowledge base to generate corresponding semantic feature vectors, denoted as . ,in, Let be the feature vector corresponding to the j-th knowledge entry in the knowledge base, and m be the total number of entries in the knowledge base; S22. Denote the obtained standardized publicity demand structured dataset as D, where D contains n dimensions of demand feature fields. The required global feature vector is generated by encoding the global features of D using a domain-pretrained BERT encoding model. Calculate the global feature vector required Feature vectors of each entry in the knowledge base The cosine similarity is used to filter out knowledge items with matching scores higher than the preset value to generate a prior constraint feature set; S23. Generate prompt words based on structured datasets and prior constraint feature sets; S24. Input the prompt word "Prompt" into the pre-fine-tuned script to generate a large model and generate an initial candidate script set. The generation stage uses a beam search algorithm, sets the beam width to N, and generates N initial candidate scripts to form a candidate script set. For each candidate script, a conditional probability is calculated, and the candidate script with the highest conditional probability is selected as the baseline script; a standardized format encapsulation is performed to generate a video script; the conditional probability is the candidate script. The model generation probability under the prompt word "Prompt" constraint; In step S4, when generating a single scene, the main action trajectory of the scene is used as a temporal hard constraint. An inter-frame motion optical flow field is generated through an optical flow prediction model. The specific implementation of constraining the large video model to generate consecutive frames of the scene is as follows: S41. Decompose the main body motion trajectory in the storyboard into a continuous set of time nodes according to the time anchor points, bind the main body posture coordinates and motion velocity to each time node, and construct the motion trajectory constraint matrix M. S42. Input the core elements of the scene and the motion trajectory temporal constraint matrix M from the storyboard into the optical flow prediction model. Initialize the inter-frame initial optical flow field guided by the trajectory gradient, and generate the target motion optical flow field through constraint iteration optimization. ,in, Let be the initial optical flow field tensor. The coordinates of any pixel point in the plane of each frame image. The trajectory guidance weights have values ​​ranging from 0 to 1. To calculate the spatial gradient of matrix M, the subject motion trend information in the motion trajectory temporal constraint matrix M is transformed into a correction value for the optical flow field. S43. Decompose the target motion optical flow field into frame-by-frame motion constraint vector groups, and convert the motion constraint vector groups into inter-frame motion hard constraint instructions that can be parsed by the large video model, thereby locking the motion direction and motion speed of each frame pixel. S44. The motion hard constraint command is generated as an immutable forced condition between frames, driving the model to strictly follow the pixel motion trajectory defined by the optical flow field to generate continuous video frames of the storyboard frame by frame.

2. The intelligent design method for video promotional materials based on a large model and knowledge base according to claim 1, characterized in that, The specific implementation of the preprocessing operation to decompose natural language requirements into a standardized structured dataset of promotional requirements includes: S11. Perform a full character scan on the natural language requirement text for video promotional material design input by the user, and remove invisible characters, special symbols, invalid spaces, and duplicate or redundant characters to obtain the initial screening requirement text. S12. Perform semantic filtering on the initial screened text to remove irrelevant modifiers and obtain the semantically filtered text. S13. Finally, intent recognition is performed, and the semantically filtered text is compared with standardized professional text in the preset industry knowledge base. The matching degree is calculated and a standardized structured dataset is generated.

3. The intelligent design method for video promotional materials based on a large model and knowledge base according to claim 2, characterized in that, The specific implementation of semantic filtering of the initially screened text and removal of irrelevant modifiers to obtain the semantically filtered text in step S12 is as follows: S121. Input the initial screened text into the semantic field construction module of the large model, generate a global semantic field with whole sentences as the carrier, map all words or phrases in the text to semantic nodes in the semantic field, and form a continuous semantic field distribution. S122. Through the gradient calculation of the semantic field of the large model, the gradient value in the semantic field is automatically identified. Nodes with gradient values ​​higher than the set gradient threshold are marked as core semantic nodes, and the remaining nodes are modifier nodes. S123. Calculate the semantic field association gradient K between the modifier node and the core semantic node. The semantic field association gradient K represents the absolute gradient value of the mean gradient of the core semantic node and the mean gradient of the modifier node, and adaptively generate a gradient filtering threshold. ,in, These are the upper and lower bounds of the semantic field gradient, respectively. We traverse all modifier nodes, remove irrelevant modifiers whose associated gradients are less than the adaptive gradient filtering threshold, and output the semantically filtered text without redundancy.

4. The intelligent design method for video promotional materials based on a large model and knowledge base according to claim 2, characterized in that, In step S13, intent recognition is performed, and the semantically filtered text is compared with standardized professional text in a preset industry knowledge base. The matching degree is calculated, and a standardized structured dataset is finally generated. The specific implementation of this step is as follows: S131. Perform semantic segmentation on the semantically filtered text, breaking it down into several minimum intent semantic units; S132. Input several minimum intent semantic units into the domain adaptive pre-trained BERT model to generate intent semantic unit feature vectors; S133. Construct a standardized semantic text library. Perform cosine similarity matching between the intention semantic unit feature vector and the full set of standardized semantic feature vectors in the standardized professional text library. Calculate the matching score and map all semantically filtered texts to the standardized semantics with the highest matching score to obtain a standardized structured dataset.

5. The intelligent design method for video promotional materials based on a large model and knowledge base according to claim 1, characterized in that, The specific implementation of step S3, which involves performing a segmentation operation on the video script, is as follows: S31. Perform global feature extraction on the video script, extracting the semantic features and scene features of each set time period in the video script; S32. Calculate the correlation between semantic features and scene features within two consecutive set time periods, and obtain the normalized correlation value by weighted summation; S33. Preset correlation threshold: When the correlation value of two consecutive time periods is lower than the correlation threshold, it is automatically determined as the corresponding segmentation benchmark point to perform the segmentation operation, and the initial time of each storyboard is marked as the time sequence anchor point.

6. The intelligent design method for video promotional materials based on a large model and knowledge base according to claim 1, characterized in that, In step S5, anomaly verification is performed on the initial video sequence. The extraction of abnormal segments is achieved by breaking down the initial video sequence into continuous image frames, performing factual logic verification on the continuous image frames, identifying content that violates physical rules and propaganda factual standards, and determining it as an abnormal segment.

Images