A natural language driven video generation method based on intention deconstruction

Abstract

This application provides a natural language-driven video generation method based on intent deconstruction, relating to the field of video generation technology, including the following steps: receiving a user's natural language intent description, thereby obtaining a structured event sequence containing multiple event nodes arranged in chronological order, each event node containing structured attribute labels for scene, subject, and behavior, and for logically related event node pairs, they are marked as at least one of temporal relationship or causal dependency relationship; based on the structured event sequence and the logical relationship between its nodes, generating a multi-condition control signal group, integrating it with a video generation model, and driving the video generation model to generate a video stream corresponding to the structured event sequence; wherein, the first type of condition signal is encoded as a text prompt embedding in the denoising process of the video generation model, and the second type of condition signal is encoded as a spatiotemporal attention mask or motion dynamics prior vector acting on the latent feature space of the video generation model.

Description

Technical Field

[0001] This invention belongs to the field of video generation technology, specifically relating to a natural language-driven video generation method based on intent deconstruction. Background Technology

[0002] In recent years, with breakthroughs in generative artificial intelligence such as diffusion models, natural language-driven video generation technology has made significant progress. Existing technical solutions mainly follow two paradigms: one is an end-to-end text-video generation model, which directly learns the mapping from text descriptions to video pixels through large-scale video-text pair training; the other is a temporal extension method based on image generation models, which first generates keyframes from text, and then generates intermediate frames through interpolation or prediction.

[0003] However, these mainstream methods reveal fundamental limitations when dealing with complex narrative intentions involving multiple events and internal logic. The core problem lies in the fact that existing models essentially learn the statistical correlation between text labels and visual content, lacking an explicit understanding and structured control of the temporal logic and causal dependencies within narrative intentions. This leads to the following defects in the generated videos: logical inconsistencies: the order of events may contradict the description, or causality may be reversed (e.g., "falling down" occurs before "slipping on a banana peel"); inconsistencies between subject and scene: the same subject may disappear, mutate, or have its attributes shift abruptly in the middle of the video; scene transitions are abrupt and do not conform to the development of events; poor action coherence: actions lack physical continuity on the timeline, appearing as a series of disjointed images. These defects make it difficult for existing technologies to generate long video content that conforms to human narrative cognition, is logically rigorous, and visually smooth, severely limiting their application in scenarios requiring rigorous logical expression, such as film and television-assisted creation, interactive storytelling, and educational simulations. Summary of the Invention

[0004] In view of the above-mentioned defects or deficiencies in the prior art, a natural language-driven video generation method based on intent deconstruction is provided, including the following steps:

[0005] Receive the user's natural language intent description;

[0006] The intent description is input into the temporal-causal event parsing model, and a structured event sequence is output. The temporal-causal event parsing model is built based on a pre-trained language model. The structured event sequence contains multiple event nodes arranged in chronological order. Each event node contains at least three structured attribute labels: scene, subject, and behavior. For event node pairs that are logically related, they are marked as at least one of temporal relationship or causal dependency relationship.

[0007] Based on the structured event sequence and the logical relationship between its nodes, a multi-condition control signal group is generated. The multi-condition control signal group includes: a first type of condition signal for controlling the consistency between the content of a single frame and the event attributes, and a second type of condition signal for controlling the visual transition and logical coherence between events.

[0008] The multi-condition control signal group is integrated with the video generation model to drive the video generation model to generate a video stream corresponding to the structured event sequence; wherein the integration is achieved by: encoding the first type of condition signal as a text prompt embedding of the video generation model in the denoising process, and encoding the second type of condition signal as a spatiotemporal attention mask or motion dynamics prior vector acting on the latent feature space of the video generation model.

[0009] According to the technical solution provided in this application, the time-series-causal event analysis model is trained through the following steps:

[0010] Construct a training dataset in which each sample includes a natural language narrative text and a corresponding manually annotated structured event sequence and logical relationship graph;

[0011] Using the narrative text as input and the joint representation of the structured event sequence and logical relationship graph as the supervision target, the pre-trained language model is fine-tuned to obtain the temporal-causal event parsing model.

[0012] According to the technical solution provided in this application, the generation of the multi-condition control signal group includes the following steps:

[0013] For each event node in the structured event sequence, its scene, subject, and behavior attribute tags are extracted and combined into a corresponding text prompt, which serves as the first type of condition signal for that event node.

[0014] Based on the logical relationship between event nodes, a spatiotemporal constraint graph is constructed; for two event nodes marked as causal dependencies, the semantic similarity of their corresponding first-type conditional signals in the text embedding space is calculated, and content anchoring strength parameters are generated based on the semantic similarity.

[0015] The spatiotemporal constraint graph and the content anchoring strength parameters are compiled together into the second type of condition signal.

[0016] According to the technical solution provided in this application, encoding the second type of conditional signal into a spatiotemporal attention mask acting on the latent feature space of the video generation model includes the following steps:

[0017] The spatiotemporal constraint graph is mapped into a three-dimensional attention weight matrix, where the three dimensions correspond to the batch dimension, spatial height dimension, and spatial width dimension of the video frame, respectively.

[0018] In a specified cross-attention layer of the U-Net decoder of the video generation model, the attention weight matrix is ​​introduced as an additional attention bias to enhance the semantic consistency of the image region corresponding to the causal dependent event during the denoising process, while weakening the interference of irrelevant regions.

[0019] According to the technical solution provided in this application, encoding the second type of conditional signal into a motion dynamics prior vector includes the following steps:

[0020] For consecutive event nodes with temporal relationships, extract the action verbs from their behavioral attribute tags;

[0021] The action verbs are converted into corresponding motion trajectory feature vectors using a pre-trained motion dynamics encoder;

[0022] The motion trajectory feature vectors of consecutive event nodes are smoothly interpolated to form a prior continuous motion trajectory across event nodes, which is then used as the motion dynamics prior vector and input to the optical flow prediction module or motion compensation module of the video generation model.

[0023] According to the technical solution provided in this application, the generation of content anchoring strength parameters based on semantic similarity, achieved through a logical influence factor prediction network, includes the following steps:

[0024] The first type of conditional signal for each pair of event nodes with causal dependency is input into the text encoder to obtain its semantic embedding, and the semantic similarity between the two is calculated.

[0025] Simultaneously, the attribute label sets of the preceding event node and the subsequent event node are concatenated and input into the logical influence factor prediction network to predict the logical influence factor vector. The logical influence factor vector includes at least a causal necessity strength component that characterizes the causal necessity strength.

[0026] The semantic similarity and the causal necessity strength component are weighted and fused to generate the content anchoring strength parameter.

[0027] According to the technical solution provided in this application, the logical influence factor vector further includes a scene transition component that characterizes the rationality of scene transitions and an action continuity component that characterizes the intensity of action continuity.

[0028] According to the technical solution provided in this application, the method further includes the following steps:

[0029] Based on the scene transition component, action coherence component, and causal necessity intensity component in the logical influence factor vector, a multi-channel spatial modulation map is generated.

[0030] When encoding the second type of conditional signal into a spatiotemporal attention mask, the multi-channel spatial modulation map is multiplied element-wise with the attention weight matrix to finely modulate the attention bias at different spatial locations in the spatiotemporal attention mask, corresponding to different logical dimensions.

[0031] According to the technical solution provided in this application, generating a multi-channel spatial modulation map includes the following steps:

[0032] The scene transition component, action coherence component, and causal necessity intensity component are subjected to time-domain alignment and smoothing filtering.

[0033] The time domain alignment refers to mapping each component value to the time axis of video generation based on the time positions of the preceding event node and the subsequent event node in the structured event sequence, and ensuring that each component value decays to a preset baseline level during time intervals without event coverage.

[0034] The smoothing filtering process is used to eliminate numerical abrupt changes in the modulation map in the time dimension caused by event node switching, and to generate a multi-channel spatial modulation map that evolves continuously in time.

[0035] According to the technical solution provided in this application, before multiplying the multi-channel spatial modulation map element-wise with the attention weight matrix, the following steps are also included:

[0036] Based on the causal necessity intensity component in the logical influence factor vector, determine whether there is a key frame interval in the video segment corresponding to the subsequent event node where the causal effect should be significantly visualized.

[0037] If so, within the keyframe interval, based on the magnitude of the causal necessity intensity component, the intensity of the spatial channel region associated with the subsequent event node and key object in the multi-channel spatial modulation map is increased to generate an enhanced spatial modulation map.

[0038] Compared with the prior art, the beneficial effects of this application are as follows:

[0039] First, it achieves precise deconstruction and reliable execution of complex narrative intentions: Through a chronological-causal event analysis model, ambiguous natural language descriptions are deconstructed into a structured sequence of events containing scenes, subjects, behaviors, and clear logical relationships (chronological / causal). This is equivalent to providing the generation process with a precise visual script, fundamentally eliminating the problems of disordered event order and reversed causality, and ensuring that the generated video is highly consistent with the user's intentions in terms of narrative logic.

[0040] Second, it provides fine-grained, multi-dimensional, decoupled generation control: by generating first-type and second-type conditional signals, content generation control is separated from logical coherence control. The first-type signal ensures that the static content of each frame strictly conforms to the attribute definition of the event node; the second-type signal is specifically responsible for regulating the visual transitions and dynamic coherence between events. This decoupled control architecture allows the model to simultaneously consider the accuracy of single-frame images and the smoothness and logic of cross-frame sequences.

[0041] Third, it achieves efficient and deep integration with advanced video generation models: the integration method is highly targeted and effective. The first type of signal is encoded as a textual cue embedding, directly leveraging the powerful text understanding capabilities of existing diffusion models to control content; the second type of signal is encoded as a spatiotemporal attention mask or a motion dynamics prior vector. The former strengthens the semantic consistency of causally related regions by influencing the spatial attention mechanism within the model, while the latter ensures the physical rationality of actions by providing explicit motion trajectory priors. This integration method does not involve coarse-grained adjustments outside the model, but rather delves into the model's latent feature space and key modules (such as the U-Net attention layer and optical flow module) for targeted intervention, enabling the logical control signals to be executed efficiently and accurately by the model.

[0042] Fourth, it significantly improves the overall quality and usability of generated videos: The generated videos are not only visually clear and realistic, but more importantly, they possess excellent logical coherence, subject consistency, and action continuity in their narrative. This enables the generated videos to meet more professional and complex application needs, such as generating logically sound short stories, instructional demonstration videos, and product function simulation animations, greatly expanding the practical boundaries and application value of natural language-driven video generation technology. Attached Figure Description

[0043] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0044] Figure 1 A flowchart illustrating the steps of the intention-based natural language-driven video generation method provided in this application. Detailed Implementation

[0045] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0046] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0047] As mentioned in the background section regarding the technical problems, this application proposes a natural language-driven video generation method based on intent deconstruction, such as... Figure 1 As shown, it includes the following steps:

[0048] S1. Receive the user's natural language intent description;

[0049] S2. Input the intent description into the temporal-causal event parsing model and output a structured event sequence. The temporal-causal event parsing model is built based on a pre-trained language model. The structured event sequence contains multiple event nodes arranged in chronological order. Each event node contains at least three structured attribute labels: scene, subject, and behavior. Event node pairs that are logically related are marked as at least one of temporal relationship or causal dependency relationship.

[0050] S3. Based on the structured event sequence and the logical relationship between its nodes, generate a multi-condition control signal group, which includes: a first type of condition signal for controlling the consistency between the content of a single frame and the event attributes, and a second type of condition signal for controlling the visual transition and logical coherence between events.

[0051] S4. Integrate the multi-condition control signal group with the video generation model to drive the video generation model to generate a video stream corresponding to the structured event sequence; wherein, the integration is achieved by: encoding the first type of condition signal as a text prompt embedding of the video generation model in the denoising process, and encoding the second type of condition signal as a spatiotemporal attention mask or motion dynamics prior vector acting on the latent feature space of the video generation model.

[0052] First, a user's natural language intent description refers to a piece of text input by the user that expresses a complete story or sequence of events, such as "A little boy was playing soccer in the park. He accidentally kicked the ball into a tree, and then he tried to jump up and get it down." The temporal-causal event parsing model is a specially trained natural language processing model that takes the aforementioned intent description text as input and outputs a structured representation. This model is built upon a pre-trained language model (e.g., BERT, GPT, or T5 series models) and fine-tuned to enable it to identify and extract structured information from narrative text.

[0053] The structured event sequence output by the model is the core intermediate representation of the entire method. It is a list where elements are called event nodes. Each event node corresponds to a sub-event in the user description and is arranged in chronological order of occurrence. Each event node is not a simple text fragment but is parsed into at least three explicit structured attribute labels: scene (describing the location or environment where the event occurs, such as "park"), subject (describing the main entity performing the action, such as "little boy"), and action (describing the action performed by the subject, such as "kicking a ball"). Furthermore, the model analyzes the relationships between event nodes. For event node pairs that are directly related in the narrative (i.e., logically connected event node pairs), the model labels the type of relationship between them, primarily temporal relationships (indicating that one event occurs after another) or causal dependencies (indicating that one event is the cause of another). Not all event nodes must have labeled relationships; the model only labels pairs with explicit logical connections.

[0054] Next, the multi-condition control signal group is a set of parameters generated based on the above structured event sequence to guide the video generation model. It is clearly divided into two categories: the first category of condition signals is responsible for ensuring that the content of each static frame in the video is consistent with the attributes (scene, subject, behavior) of the corresponding event node; the second category of condition signals is responsible for controlling how the video segments corresponding to different events transition, ensuring visual smoothness and logical coherence of events.

[0055] Finally, the video generation model refers to existing video generation models based on the denoising diffusion principle, such as StableVideo Diffusion or models with similar architectures. The integration of this method involves adapting the generated control signal set into the internal working mechanism of such models. Specifically, the first type of conditional signal is converted into textual cue embeddings that the model can understand, serving as conditional input in each step of the denoising generation process; simultaneously, the second type of conditional signal is encoded into one of two possible forms, directly intervening in the model's internal computation. One is a spatiotemporal attention mask, a matrix that acts on the attention layer in the U-Net network to change the model's attention level when generating different spatial regions and time frames; the other is a motion dynamics prior vector, a feature vector that is input into the model's motion prediction or compensation module, providing high-level guidance for the motion of objects in the video.

[0056] In implementation, the system first receives text input from the user. Then, a pre-trained temporal-causal event parsing model is invoked to parse the text into a structured sequence of events. Next, a processing program is written to traverse this sequence: for each event node, its attribute labels are combined into a short sentence (e.g., "A little boy is playing football in the park") as a first-type signal; simultaneously, the logical relationships between the labels of the nodes are analyzed, a graph structure representing the constraints between events is constructed, and for causal dependency pairs, an intensity parameter is generated by calculating the semantic vector similarity of their first-type signal sentences and combining it with a prediction network (detailed later). Finally, the graph and intensity parameter are compiled into an intermediate representation of the second-type signal. Finally, when the video generation model is invoked, the list of first-type signal sentences is provided to the model in chronological order as text conditions, while the intermediate representation of the second-type signal is converted into an attention mask or motion prior vector according to the selected method and injected into the specified components of the model to drive the model to generate the final video stream.

[0057] This implementation addresses the problems of existing end-to-end text-to-video generation technologies, such as difficulty in handling complex narrative logic, and the tendency to produce disordered event sequences and inconsistent content. Its technical principle lies in introducing a structured intermediate representation layer (structured event sequence) to deconstruct ambiguous intentions into clear, discrete event units with logical relationships. Furthermore, targeted control signals are designed for these units and relationships, and precisely applied using the video generation model's interventionist interfaces (text conditions, attention mechanisms, motion modules). This reliably maps high-level narrative logic to a low-level visual generation process, achieving logically controllable video generation.

[0058] In a preferred embodiment, the time-series-causal event analysis model is trained through the following steps:

[0059] Construct a training dataset in which each sample includes a natural language narrative text and a corresponding manually annotated structured event sequence and logical relationship graph;

[0060] Using the narrative text as input and the joint representation of the structured event sequence and logical relationship graph as the supervision target, the pre-trained language model is fine-tuned to obtain the temporal-causal event parsing model.

[0061] First, a training dataset needs to be constructed. Each sample in the dataset consists of two parts: the first part is a natural language narrative text, that is, a text describing a series of events, such as a short story, a script, or an event narrative. The second part is a manually labeled corresponding structured event sequence and logical relationship graph, which is a supervised label created by professionals based on the aforementioned narrative text. The labeling process includes: 1) dividing the text into multiple event nodes; 2) labeling each node with attribute labels such as scene, subject, and behavior; 3) arranging all nodes in chronological order; 4) determining which node pairs have logical connections and labeling them as either temporal relationships or causal dependencies, thus forming a logical relationship graph. This relationship graph can be represented in the form of an adjacency matrix or an edge list, with the relationship type indicated on the edges.

[0062] Then, the narrative text is used as input, and the joint representation of the structured event sequence and the logical relationship graph is used as the supervision objective. Here, the joint representation refers to encoding the structured event sequence (which can be viewed as a sequence of attribute labels) and the logical relationship graph (the structure of nodes and edges) into a unified format that the model can learn. A common implementation is to represent the structured event sequence as a special structured text (e.g., using XML or a format with custom delimiters), and the logical relationship graph as a flattened vector of an adjacency matrix or a serialized list of edges, then concatenate the two as the objective. Another more advanced approach is to use graph neural networks to encode the relationship graph as features and then fuse it with the sequence representation.

[0063] Next, the pre-trained language model is fine-tuned using the constructed dataset. The specific steps are as follows: Select a large-scale pre-trained language model (such as BERT or T5) as the base model. Input the narrative text of each sample in the dataset into the model. Design the model's output layer and loss function so that it can simultaneously predict the structure of the text and logical relation graph of the structured event sequence. For example, the task can be constructed as a text-to-text generation task, allowing the model to directly output the aforementioned joint representation text string; alternatively, it can be constructed as a multi-task learning task, using different output heads to predict the node attribute sequence and relation matrix respectively. During training, the labeled joint representation is used as the ground truth. By calculating cross-entropy loss or mean squared error loss, and using the backpropagation algorithm to update the model parameters, the model learns the mapping rules from narrative text to structured representation. After sufficient iterations of training, a converged model is obtained, which is the desired temporal-causal event parsing model.

[0064] In a preferred embodiment, generating the multi-condition control signal group includes the following steps:

[0065] For each event node in the structured event sequence, its scene, subject, and behavior attribute tags are extracted and combined into a corresponding text prompt, which serves as the first type of condition signal for that event node.

[0066] Based on the logical relationship between event nodes, a spatiotemporal constraint graph is constructed; for two event nodes marked as causal dependencies, the semantic similarity of their corresponding first-type conditional signals in the text embedding space is calculated, and content anchoring strength parameters are generated based on the semantic similarity.

[0067] The spatiotemporal constraint graph and the content anchoring strength parameters are compiled together into the second type of condition signal.

[0068] The first step is to generate the first type of conditional signal. Specifically, this involves iterating through each event node in the structured event sequence. For the current node, the text values ​​of the three attribute labels (scene, subject, and behavior) are read from its data structure. Then, these labels are combined into a complete, descriptive natural language sentence according to a preset syntactic template. For example, if the labels are "park," "little boy," and "playing football," they can be combined into "A little boy is playing football in the park." This generated sentence is the first type of conditional signal corresponding to that event node. After all event nodes have been processed, a list of text prompts consistent with the order of the event sequence is obtained.

[0069] The second step involves generating intermediate components for the second type of conditional signals. First, a spatiotemporal constraint graph is constructed based on the logical relationships between event nodes. This spatiotemporal constraint graph is an abstract data structure used to characterize how the correlations between events affect the spatiotemporal characteristics of the video. In implementation, a graph with event nodes as nodes can be constructed. For each pair of nodes labeled with a logical relationship (temporal or causal) in the structured event sequence, a directed edge connecting them is added to the graph. Each edge can be accompanied by a basic weight or type identifier. This graph structure constitutes the initial constraints on the temporal order and spatial correlations in the video.

[0070] Secondly, further computation is needed for the two event nodes marked as causal dependencies. The first-class conditional signals corresponding to each node (i.e., the two descriptive sentences generated in the previous step) are taken. Using a pre-trained text encoder (e.g., the Sentence-BERT or CLIP text encoder), these two sentences are converted into high-dimensional semantic embedding vectors. Then, the cosine similarity or the reciprocal of the Euclidean distance between these two vectors is calculated to obtain a semantic similarity scalar value. This value reflects the closeness of the two events in terms of semantic content. Subsequently, a content anchoring strength parameter needs to be generated based on the semantic similarity. This means that semantic similarity needs to be used as an important input, and may also need to be combined with other information (such as subsequent logical influence factors), through a function or a small network, to calculate a final strength parameter. This parameter will be used to indicate the extent to which the visual content of this pair of causal events needs to be anchored or maintained during video generation to avoid unreasonable abrupt changes in the subject or scene.

[0071] The third step involves compiling the spatiotemporal constraint graph and the content anchoring strength parameters together into a second type of conditional signal. This refers to integrating the graph structure obtained above and the strength parameters for causal edges into a unified data object that can be used by subsequent encoding operations (such as attention mask generation). For example, in implementation, an enhanced graph data structure can be created where each edge not only contains the relation type, but also, for causal dependency edges, the calculated content anchoring strength parameter value.

[0072] This implementation achieves a reliable conversion from structured events to specific control parameters, particularly by quantifying visual consistency requirements through calculating the semantic similarity between causal events. The technical principle is to map logical relationships (especially causal dependencies) into computable constraints (graphs) and strength parameters, thereby providing quantifiable guidance for the subsequent generation process and making the abstract requirement of logical coherence operable.

[0073] In a preferred embodiment, encoding the second type of conditional signal into a spatiotemporal attention mask acting on the latent feature space of the video generation model includes the following steps:

[0074] The spatiotemporal constraint graph is mapped into a three-dimensional attention weight matrix, where the three dimensions correspond to the batch dimension, spatial height dimension, and spatial width dimension of the video frame, respectively.

[0075] In a specified cross-attention layer of the U-Net decoder of the video generation model, the attention weight matrix is ​​introduced as an additional attention bias to enhance the semantic consistency of the image region corresponding to the causal dependent event during the denoising process, while weakening the interference of irrelevant regions.

[0076] First, the spatiotemporal constraint graph needs to be mapped into a three-dimensional attention weight matrix. This spatiotemporal constraint graph is the previously constructed graph, possibly with attached intensity parameters. The goal of the mapping is to generate a three-dimensional tensor of dimensions (B, H, W) as the attention weight matrix. Here, B corresponds to the batch dimension, which in video generation typically corresponds to a sequence (or fragment) of time frames; H and W correspond to the spatial height and spatial width dimensions, respectively, representing the spatial raster of the final generated video frames. The specific implementation is as follows: Assume we want to generate a video of length T frames. We can divide the timeline into T intervals, with each event node occupying a portion of a continuous frame interval according to its order in the structured sequence. For each edge in the spatiotemporal constraint graph (indicating a relationship between two event nodes A and B), its influence range in the temporal dimension can be set to cover all or part of the frame intervals corresponding to nodes A and B. In the spatial dimension, we need to predefine or estimate the main image region of interest for each event node (e.g., a general region can be associated with it through subject labels). Then, for each frame (a slice in the B-dimensional dimension) within the influence time range in the weight matrix, the weight value is set to a high positive number (e.g., 1.0 or an accompanying intensity parameter value) within the region of interest corresponding to the associated event node in space, and a low value (e.g., 0 or a negative bias) within the region of interest. For causal dependencies, the intensity parameter can be used to modulate the specific value of this high weight. Finally, a three-dimensional matrix is ​​obtained, where the values ​​indicate which spatiotemporal locations of the feature map the model should focus on (positive bias) or ignore (negative bias or zero bias) at each step of the denoising process.

[0077] Then, in a specified cross-attention layer of the U-Net decoder in the video generation model, the attention weight matrix is ​​introduced as an additional attention bias. A typical video diffusion model, U-Net, contains multiple cross-attention layers for fusing textual conditions. During implementation, one or more layers need to be selected for intervention. When the model's forward computation reaches this specified layer, after performing standard query (Q), key (K), and value (V) calculations and obtaining the attention score matrix, instead of directly performing softmax normalization, the previously generated 3D attention weight matrix (which may require appropriate dimensionality broadcasting or scaling to fit the shape of the attention scores) is added as an additional attention bias to the original attention score matrix. This operation is equivalent to modifying the distribution of the attention weights. Afterward, softmax and a weighted summation with V are performed.

[0078] The aim is to enhance the semantic consistency of image regions corresponding to causally dependent events during the denoising process, while weakening the interference from irrelevant regions. The technical principle is as follows: by introducing this external bias, when generating images of two causally related events, the model is guided to pay more attention to the features of spatial regions related to the consequence in the preceding event image, and when generating the consequence event image, it tends to reuse or maintain consistent feature representations, thereby visually strengthening the causal connection. Simultaneously, for unrelated events or regions in the image, the bias reduces the attention they receive, lowering the possibility of content drift caused by irrelevant information interference.

[0079] This implementation provides a very specific and effective mechanism that directly transforms the logical relationships between events into interventions in the core attention mechanism of the generative model, thereby achieving pixel-level spatially controllable enhancement of logical consistency. Its technical principle utilizes the plasticity of the attention mechanism in the diffusion model, purposefully guiding the feature aggregation process by adding external spatial biases. This is a method for targeted editing of the model's internal representation.

[0080] In a preferred embodiment, encoding the second type of conditional signal into a motion dynamics prior vector includes the following steps:

[0081] For consecutive event nodes with temporal relationships, extract the action verbs from their behavioral attribute tags;

[0082] The action verbs are converted into corresponding motion trajectory feature vectors using a pre-trained motion dynamics encoder;

[0083] The motion trajectory feature vectors of consecutive event nodes are smoothly interpolated to form a prior continuous motion trajectory across event nodes, which is then used as the motion dynamics prior vector and input to the optical flow prediction module or motion compensation module of the video generation model.

[0084] First, the execution step extracts the action verbs from the behavioral attribute labels of consecutive event nodes with a temporal relationship. During implementation, the program traverses the structured event sequence. When two event nodes are consecutive in the sequence (i.e., adjacent) and their logical relationship is marked as temporal, they are considered a pair of consecutive nodes. For each node, the core action verb is extracted from the text of its behavioral attribute label. For example, if the behavioral label is "running quickly," then "running" is extracted; if it is "picking up a cup," then "picking up" is extracted. This can be achieved using dependency parsing tools or simple verb lexicon matching.

[0085] Secondly, a pre-trained motion dynamics encoder is implemented to convert the action verbs into corresponding motion trajectory feature vectors. Here, the pre-trained motion dynamics encoder is a standalone machine learning model whose function is to map the text describing the action into a vector representing the typical motion pattern of that action. This encoder can be trained on large-scale motion capture data (such as Human3.6M) and action description text pairs. During implementation, the extracted action verbs (such as "running" and "jumping") are input into the encoder, which outputs a fixed-dimensional feature vector. This vector can characterize multiple aspects of the action, such as the main direction of motion, velocity profile, and coordination patterns of limb joints.

[0086] The third step involves smoothing the motion trajectory feature vectors of consecutive event nodes through interpolation, forming a continuous motion trajectory prior across event nodes. Assume event node A (verb "run-up") corresponds to the first 10 frames of the video, with its motion feature vector being Va; node B (verb "jump") corresponds to frames 11 to 20, with its motion feature vector being Vb. Simply switching motion between frames 10 and 11 would result in discontinuity. Therefore, interpolation over time is necessary. One implementation method is to perform linear or non-linear (e.g., spherical linear) interpolation on Va and Vb within several frames (e.g., frames 8, 9, 10, 11, and 12) between event A and event B, calculating a transitional motion feature vector for each frame. Ultimately, a motion trajectory feature vector is assigned to each frame of the entire video, forming a continuously changing "motion trajectory prior" sequence over time. This sequence is the final form of the "motion dynamics prior vector".

[0087] Finally, the prior vector is input into the optical flow prediction module or motion compensation module of the video generation model. In diffusion-based video generation, there is often a dedicated submodule responsible for predicting inter-frame optical flow or performing motion compensation to ensure temporal smoothness. In implementation, the motion trajectory feature vector corresponding to each frame, calculated above, needs to be input as a condition into these modules. For example, in the optical flow prediction module, this vector can be concatenated with the latent features of the current frame to jointly predict the optical flow field of the next frame, thereby making the motion of the generated object more consistent with the dynamic pattern indicated by the input action verb.

[0088] This implementation addresses the problem of stiff and physically incompatible motion in existing video generation, providing explicit and coherent motion guidance for sequential events. Its technical principle lies in decoding high-level behavioral semantics (action verbs) into low-level motion pattern representations through a pre-trained model, ensuring the continuity of motion during event transitions through temporal interpolation, and finally injecting this prior knowledge into the motion-related modules of the video generation model to guide the generation of dynamically reasonable videos.

[0089] In a preferred embodiment, the generation of content anchoring strength parameters based on semantic similarity is achieved through a logical influence factor prediction network, and includes the following steps:

[0090] The first type of conditional signal for each pair of event nodes with causal dependency is input into the text encoder to obtain its semantic embedding, and the semantic similarity between the two is calculated.

[0091] Simultaneously, the attribute label sets of the preceding event node and the subsequent event node are concatenated and input into the logical influence factor prediction network to predict the logical influence factor vector. The logical influence factor vector includes at least a causal necessity strength component that characterizes the causal necessity strength.

[0092] The semantic similarity and the causal necessity strength component are weighted and fused to generate the content anchoring strength parameter.

[0093] The first step involves inputting the first-class conditional signals of each pair of event nodes with causal dependencies into a text encoder to obtain their semantic embeddings, and then calculating their semantic similarity. This is consistent with the steps for calculating semantic similarity. Specifically, the first-class conditional signals (i.e., two descriptive sentences) of the preceding and following event nodes in the causal pair are extracted. Using a fixed, pre-trained text encoder, such as the Sentence-BERT model, the two sentences are encoded into two high-dimensional vectors, E_cause and E_effect. Subsequently, the cosine similarity between these two vectors is calculated, yielding a scalar value S_semantic between -1 and 1. This value directly reflects the closeness of the two event descriptions in the general semantic space.

[0094] The second step involves concatenating the attribute label sets of preceding and subsequent event nodes in parallel, and inputting this concatenation into the logical influence factor prediction network to predict the logical influence factor vector. Here, the attribute label set refers to the text set of labels for each event node, such as scene, subject, and behavior. During implementation, all the label texts of the two nodes need to be concatenated into a long text sequence in a fixed order, for example, "Scene: Park; Subject: Boy; Behavior: Kicking a ball; Scene: Park; Subject: Boy; Behavior: Retrieving a ball". This concatenated sequence is input into the logical influence factor prediction network. This network is a specially trained neural network, whose architecture can be a small Transformer-based encoder followed by a Multilayer Perceptron (MLP) as the output head. The network's training objective is to learn to infer the strength and nature of the causal relationship between event pairs from their attribute descriptions. The network outputs a vector that includes at least a causal necessity strength component representing the strength of causal necessity. This component is a scalar value, for example, between 0 and 1, indicating the degree to which the subsequent event is necessarily caused by the preceding event.

[0095] The third step involves weighted fusion of the semantic similarity and the causal necessity strength component to generate a content anchoring strength parameter. This integrates information from two different sources. In practice, a simple fusion function can be designed. For example, let the semantic similarity be S and the causal necessity strength component be N. The final content anchoring strength parameter P can be calculated as P = α × S + β × N, where α and β are adjustable fusion weight coefficients that can be determined during model training or system tuning. A more complex approach is to let P = f(S, N), where f is a small neural network. This final generated P value will serve as a quantification metric used in subsequent steps to modulate the attention mask, determining the degree to which the content of the two causal events needs to be visually anchored.

[0096] This implementation provides a more accurate and robust method for quantifying the visual anchoring requirements of causal dependencies, combining general semantic relevance with specially learned causal logic strength. Its technical principle is dual-path information fusion: one path uses a general text encoder to capture surface semantic relevance; the other path uses a specially trained network to deeply understand the logical implications between event attributes. The two paths are complementary and weighted to derive a more reliable guidance strength parameter, overcoming the potential misjudgments that may arise from relying solely on semantic similarity (e.g., semantically similar but without logical causality, or with strong causality but significant differences in semantic description).

[0097] In a preferred embodiment, the logical influence factor vector further includes a scene transition component that characterizes the rationality of scene transitions and an action continuity component that characterizes the intensity of action continuity.

[0098] This implementation directly relies on the modification and extension of the logical influence factor prediction network. In the above implementation, the output layer of this network was designed to output at least one value, namely the causal necessity strength component. According to this implementation, the output layer of the network needs to be redesigned to enable it to output more components.

[0099] The specific implementation steps are as follows: First, keep the main structure of the network (such as Transformer encoder + MLP) unchanged. Then, modify its final output layer (MLP). Assume the original output layer outputs a scalar value. Now, it needs to be expanded to output three scalar values. These three output nodes correspond to:

[0100] The causal necessity intensity component (C): As mentioned above, it indicates whether a subsequent event is necessarily caused by a preceding event.

[0101] Scene transition plausibility component (S): This is a new component used to quantify whether the transition from the scene of a preceding event to the scene of a subsequent event is visually natural and plausible. For example, a sudden switch from "indoors" to "outer space" might be less plausible, while a switch from "living room" to "kitchen" would be more plausible. This component requires the network to learn the probability of such common-sense transitions from scene labels.

[0102] Action coherence intensity component (A): This is another new component used to quantify whether the actions of preceding events and subsequent events are coherent and smooth in the action sequence. For example, the coherence from "running up" to "jumping" is high, while the coherence from "sitting down" to "running" is low. This component requires the network to learn common patterns of action sequences from action verbs.

[0103] To train this network that outputs three components, the training dataset needs to be expanded. During annotation, it's necessary not only to indicate whether causal dependencies exist between event pairs, but also for annotators to score each pair of events with a temporal or causal relationship based on two dimensions: scene transition and action coherence (e.g., 0-1 points). These scores will serve as supervision signals, used along with causal necessity labels, to train the network. The loss function can be designed as the sum of the mean squared error losses of the three components.

[0104] During inference, for a pair of input event attribute labels, the network will simultaneously output the values ​​of these three components, which together constitute a three-dimensional "logical influence factor vector" [C, S, A]. This vector contains more comprehensive logical relationship information than a single component.

[0105] This implementation refines and enriches the representation of logical influence, expanding from a single causal necessity to encompass scene transitions and action coherence, thereby providing multi-dimensional and more refined logical control for video generation. Its technical principle is to expand the output space and supervisory signals of the neural network, forcing the model to explicitly model and output the correlation strength between events across these different logical dimensions during the learning process. This allows the generated control signals to simultaneously consider multiple aspects such as content causality, environmental transitions, and action continuity.

[0106] In a preferred embodiment, the method further includes the following steps:

[0107] Based on the scene transition component, action coherence component, and causal necessity intensity component in the logical influence factor vector, a multi-channel spatial modulation map is generated.

[0108] When encoding the second type of conditional signal into a spatiotemporal attention mask, the multi-channel spatial modulation map is multiplied element-wise with the attention weight matrix to finely modulate the attention bias at different spatial locations in the spatiotemporal attention mask, corresponding to different logical dimensions.

[0109] The first step involves using a multi-channel spatial modulation map, which is a three-dimensional tensor with dimensions (C, H, W), where C is the number of channels, and H and W are the spatial height and width. In this implementation, C can be set to 3, with each channel specifically corresponding to a component in the logical influence factor vector. The specific generation method is as follows: For a pair of logically related event nodes, a time window is first determined based on their influence range on the video timeline (e.g., the frame interval they occupy). For each frame within this window, a corresponding (C, H, W) modulation map needs to be generated. The numerical initialization of each channel can be based on the scalar value of the corresponding component. For example, if the causal necessity strength component value is 0.8, then 0.8 is filled in the "causal necessity channel" of the entire map (or the predicted area of ​​the preceding event subject); if the scene transition component value is 0.5, then 0.5 is filled in the "scene transition channel". A more refined approach is to apply different components to different spatial regions: the causal component may focus more on the subject and directly affected object areas, while the scene transition component may act more evenly on the global background area. This allows for a rough estimation of the spatial mask using the attribute labels of event nodes (such as subject, key object), and then the component values ​​are assigned to the corresponding mask regions.

[0110] The second step involves encoding the second type of conditional signal into a spatiotemporal attention mask by performing element-wise multiplication of the multi-channel spatial modulation map with the attention weight matrix. The spatiotemporal constraint map is mapped to a three-dimensional attention weight matrix M_attn(B, H, W). In this step, the multi-channel spatial modulation map M_mod(C, H, W) generated in the previous step is first fused along the channel dimension, converting it into a single-channel modulation map matching the spatial dimension of M_attn. A simple fusion method is to perform a weighted summation of the three channels: M_fused = w1 × M_mod [causal channel] + w2 × M_mod [scene channel] + w3 × M_mod [action channel]. The weights w1, w2, and w3 can be preset or learned. Then, along the temporal dimension, M_fused is broadcast to each frame (B frames), resulting in M_fused_expanded(B, H, W). Finally, an element-wise multiplication operation is performed: M_enhanced = M_attn * M_fused_expanded. Here, * indicates that the values ​​at corresponding positions are multiplied. The result of this multiplication is that the original attention bias weights M_attn are scaled spatially by the logistic factor modulation map. In regions emphasized by the logistic factor (where the product value is larger), the attention bias is enhanced; in regions not emphasized by the logistic factor, the attention bias is weakened.

[0111] The ultimate goal is to finely modulate the attention biases corresponding to different logical dimensions at different spatial locations within the spatiotemporal attention mask. This allows the model to apply attention biases in a more nuanced and refined manner, rather than uniformly, based on the needs of different logical dimensions such as causality, scene, and action, in different areas of the image.

[0112] This implementation transforms the results of multidimensional logical relationship analysis into pixel-level, channel-separated spatial control signals, and deeply integrates them with the attention mechanism, achieving unprecedented fine-grained logical guidance for the video generation process. Its technical principle is to establish a mapping from logical components to spatial channels, and use this mapping result as coefficients to dynamically modulate the intensity of the attention bias, thereby achieving decoupling of logical control and interpretable, fine-grained intervention in the spatial dimension.

[0113] In a preferred embodiment, generating a multi-channel spatial modulation map includes the following steps:

[0114] The scene transition component, action coherence component, and causal necessity intensity component are subjected to time-domain alignment and smoothing filtering.

[0115] The time domain alignment refers to mapping each component value to the time axis of video generation based on the time positions of the preceding event node and the subsequent event node in the structured event sequence, and ensuring that each component value decays to a preset baseline level during time intervals without event coverage.

[0116] The smoothing filtering process is used to eliminate numerical abrupt changes in the modulation map in the time dimension caused by event node switching, and to generate a multi-channel spatial modulation map that evolves continuously in time.

[0117] First, temporal alignment and smoothing filtering are required for the scene transition component, action coherence component, and causal necessity strength component. Temporal alignment is the core step. In implementation, assume we have a causal event pair: preceding event node A (e.g., "lifting the ball") and subsequent event node B (e.g., "throwing the ball"), with corresponding logical influence factor vectors [C, S, A], representing causal necessity, scene transition rationality, and action coherence strength, respectively. In video generation, event A and event B each occupy a continuous time interval. Assuming the total video duration is T frames, event A occupies frames 1 to t1, and event B occupies frames t2 to T (t2>t1), with a possible transition interval in between. The goal of temporal alignment is to expand these three discrete component scalar values ​​into three time-series signals, each of length T.

[0118] The specific implementation is as follows: Create a one-dimensional array of length T for each component, initially setting all values ​​to a preset baseline level. This baseline level is a small value close to zero (e.g., 0.1), representing that during periods without explicit event coverage, logical control should be minimized, allowing the generative model to follow its default dynamics. Then, map the component values:

[0119] For the causal necessity strength component C, since it relates to two events, its influence should primarily cover the entire causal chain from the occurrence of event A to the end of event B. Therefore, a weight distribution can be designed on the interval [1, T]. For example, a higher weight can be assigned to the core interval [1, t1] of event A, a higher or gradually changing weight can be maintained in the transition interval [t1+1, t2-1], and a certain weight can also be assigned to the interval [t2, T] of event B to maintain the visual anchoring of the result.

[0120] For the scene transition rationality component S, its influence mainly applies to the transition interval where the scene changes. Therefore, the array value can be set to S near the transition frame from the end of event A to the beginning of event B (e.g., [t1-2, t2+2]), while keeping the baseline level in other areas.

[0121] For the action coherence intensity component A, its influence should cover the entire period involved in the action execution. Therefore, the array value is set to A within the action execution interval of events A and B (which may need to be estimated based on the action verb).

[0122] Next, a smoothing filter is performed. After the alignment operation described above, each component's time series may exhibit abrupt changes at event boundaries (e.g., a sudden jump from the baseline level of 0.1 to a component value of 0.9). These abrupt changes, propagated to the spatial modulation map and affecting the attention mechanism, can cause flickering or discontinuity in the generated video at corresponding frames. Therefore, a digital filter is needed to smooth these three time series. One specific implementation is to use one-dimensional Gaussian filtering or mean filtering. For example, using a sliding window (e.g., a window size of 5 frames), for each point in the sequence, the original value is replaced with the weighted average of its surrounding points (Gaussian weights). This results in a smooth, sloping transition at event boundaries, rather than a sharp, abrupt change. After processing, we obtain three smooth time series, each of length T, representing the modulation intensity of causal necessity, scene transition rationality, and action coherence at each frame.

[0123] Finally, these three smoothed time series are combined with spatial region information to generate a multi-channel spatial modulation map that evolves continuously in time. Specifically, for the i-th frame of the video (1 ≤ i ≤ T), the smoothed intensity values ​​Ci', Si', and Ai' of the three components in that frame are extracted. Then, based on the spatial mask estimated according to the event node attributes (such as the main region and the background region), these intensity values ​​are filled or weighted to the corresponding spatial positions to form the (C, H, W) three-channel modulation map of the i-th frame. The modulation maps of all T frames are combined to form a four-dimensional tensor (T, C, H, W) that evolves in the time dimension, which is the final temporally continuous spatial modulation map.

[0124] In a preferred embodiment, before performing element-wise multiplication of the multi-channel spatial modulation map with the attention weight matrix, the method further includes the following step:

[0125] Based on the causal necessity intensity component in the logical influence factor vector, determine whether there is a key frame interval in the video segment corresponding to the subsequent event node where the causal effect should be significantly visualized.

[0126] If so, within the keyframe interval, based on the magnitude of the causal necessity intensity component, the intensity of the spatial channel region associated with the subsequent event node and key object in the multi-channel spatial modulation map is increased to generate an enhanced spatial modulation map.

[0127] This step occurs "before the element-wise multiplication of the multi-channel spatial modulation map with the attention weight matrix". Its first sub-step is: based on the causal necessity strength component in the logical influence factor vector, determine the keyframe interval in the video segment corresponding to the subsequent event node where the causal effect should be significantly visualized. In implementation, the program first reads the value of the causal necessity strength component C. A threshold is set, for example, 0.5. If C is less than this threshold, the causality is considered weak, requiring no special enhancement, and therefore no subsequent enhancement steps are performed; the original spatial modulation map is used directly. If C is greater than or equal to the threshold, enhancement is deemed necessary. Next, the keyframe interval needs to be determined. This is typically located at the beginning of the time interval corresponding to the subsequent event node, as this is the moment when the causal effect first manifests and is most impactful. A specific implementation algorithm is: assuming the subsequent event node B occupies the [t_start, t_end]th frame of the video. The keyframe interval can be defined as [t_start, t_start + K], where K is a frame number dynamically calculated based on the C value. For example, K = ceil(C * M), where M is the preset maximum number of enhanced frames (e.g., 10 frames). In this way, the stronger the causal necessity (the closer C is to 1), the longer the keyframe interval, and the more lasting the enhancement effect.

[0128] The second sub-step is as follows: Within the keyframe interval, based on the magnitude of the causal necessity intensity component, the intensity of the spatial channel regions associated with the subject and key objects of the subsequent event node in the multi-channel spatial modulation map is enhanced to generate an enhanced spatial modulation map. In implementation, it is first necessary to identify the "spatial channel regions associated with the subject and key objects" of the subsequent event node B. This can be inferred from the attribute labels of the event node. For example, if the description of node B is "glass shattered," its subject is "glass." When generating the spatial modulation map, the spatial location of the glass (which can be roughly estimated through object recognition in the text prompt or by using a prior object attention map) is marked as the relevant region. Then, for each frame within this keyframe interval, its multi-channel spatial modulation map (still three channels: causality, scene, and action) is manipulated. The focus is on enhancing the values ​​of the causal necessity channel (or, depending on the design, possibly the relevant region in all channels). The enhancement method can be linear scaling: new value = original value × (1 + α × C), where α is an enhancement coefficient (e.g., 0.5). In this way, the modulation intensity is significantly amplified in the relevant region. After performing this operation on the relevant regions of all frames within the keyframe interval, an enhanced spatial modulation map is obtained. For non-keyframe intervals, the original modulation map remains unchanged.

[0129] Ultimately, this enhanced spatial modulation map replaces the original multichannel spatial modulation map and is used for element-wise multiplication with the generated attention weight matrix. As a result, at critical moments when causal effects occur, the generative model receives extra attention to the subject or object region involved in the "effect," thus tending to generate more significant, detailed, and closely related visual changes in that region.

[0130] This implementation addresses the problem that existing solutions may homogenize causal control, leading to insufficient visual expressiveness and a lack of emphasis on causal effects. By dynamically identifying and enhancing key moments of causality, the generated video can, like a human director editing, provide visual close-ups at important narrative junctures, greatly enhancing the story's tension and logical clarity. Its technical principle lies in using the causal intensity scalar in logical relationships not only as a weight for spatial control but also as a basis for temporal scheduling and intensity modulation, achieving a dynamic and non-uniform mapping of logical importance to visual expressiveness.

[0131] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A natural language-driven video generation method based on intent deconstruction, characterized in that, Includes the following steps: Receive the user's natural language intent description; The intent description is input into the temporal-causal event parsing model, and a structured event sequence is output. The temporal-causal event parsing model is built based on a pre-trained language model. The structured event sequence contains multiple event nodes arranged in chronological order. Each event node contains at least three structured attribute labels: scene, subject, and behavior. For event node pairs that are logically related, they are marked as at least one of temporal relationship or causal dependency relationship. Based on the structured event sequence and the logical relationship between its nodes, a multi-condition control signal group is generated. The multi-condition control signal group includes: a first type of condition signal for controlling the consistency between the content of a single frame and the event attributes, and a second type of condition signal for controlling the visual transition and logical coherence between events. The multi-condition control signal group is integrated with the video generation model to drive the video generation model to generate a video stream corresponding to the structured event sequence; wherein, the integration is achieved by: encoding the first type of condition signal as a text prompt embedding of the video generation model in the denoising process, and encoding the second type of condition signal as a spatiotemporal attention mask or motion dynamics prior vector acting on the latent feature space of the video generation model; The generation of the multi-condition control signal group includes the following steps: For each event node in the structured event sequence, its scene, subject, and behavior attribute tags are extracted and combined into a corresponding text prompt, which serves as the first type of condition signal for that event node. Based on the logical relationship between event nodes, a spatiotemporal constraint graph is constructed; for two event nodes marked as causal dependencies, the semantic similarity of their corresponding first-type conditional signals in the text embedding space is calculated, and content anchoring strength parameters are generated based on the semantic similarity. The spatiotemporal constraint graph and the content anchoring strength parameters are compiled together into the second type of condition signal; The second type of conditional signal encoding is a spatiotemporal attention mask acting on the latent feature space of the video generation model, including the following steps: The spatiotemporal constraint graph with content anchoring strength parameters is mapped into a three-dimensional attention weight matrix, where the three dimensions correspond to the batch dimension, spatial height dimension, and spatial width dimension of the video frame, respectively. In a specified cross-attention layer of the U-Net decoder of the video generation model, the attention weight matrix is ​​introduced as an additional attention bias to enhance the semantic consistency of the image region corresponding to the causal dependent event during the denoising process, while weakening the interference of irrelevant regions. Encoding the second type of conditional signal into a motion dynamics prior vector includes the following steps: For consecutive event nodes with temporal relationships, extract the action verbs from their behavioral attribute tags; The action verbs are converted into corresponding motion trajectory feature vectors using a pre-trained motion dynamics encoder; The motion trajectory feature vectors of consecutive event nodes are smoothly interpolated to form a continuous motion trajectory prior across event nodes, which is then used as the motion dynamics prior vector and input to the optical flow prediction module or motion compensation module of the video generation model. The generation of content anchoring strength parameters based on semantic similarity is achieved through a logical influence factor prediction network, including the following steps: The first type of conditional signal for each pair of event nodes with causal dependency is input into the text encoder to obtain its semantic embedding, and the semantic similarity between the two is calculated. Simultaneously, the attribute label sets of the preceding event node and the subsequent event node are concatenated and input into the logical influence factor prediction network to predict the logical influence factor vector. The logical influence factor vector includes at least a causal necessity strength component that characterizes the causal necessity strength. The semantic similarity and the causal necessity strength component are weighted and fused to generate the content anchoring strength parameter.

2. The natural language-driven video generation method based on intent deconstruction according to claim 1, characterized in that, The time-series-causal event analysis model is trained through the following steps: Construct a training dataset in which each sample includes a natural language narrative text and a corresponding manually annotated structured event sequence and logical relationship graph; Using the narrative text as input and the joint representation of the structured event sequence and logical relationship graph as the supervision target, the pre-trained language model is fine-tuned to obtain the temporal-causal event parsing model.

3. The natural language-driven video generation method based on intent deconstruction according to claim 1, characterized in that, The logical influence factor vector also includes a scene transition component that characterizes the rationality of scene transitions and an action continuity component that characterizes the intensity of action continuity.

4. The natural language-driven video generation method based on intent deconstruction according to claim 3, characterized in that, The method also includes the following steps: Based on the scene transition component, action coherence component, and causal necessity intensity component in the logical influence factor vector, a multi-channel spatial modulation map is generated. When encoding the second type of conditional signal into a spatiotemporal attention mask, the multi-channel spatial modulation map is multiplied element-wise with the attention weight matrix to finely modulate the attention bias at different spatial locations in the spatiotemporal attention mask, corresponding to different logical dimensions.

5. The natural language-driven video generation method based on intent deconstruction according to claim 4, characterized in that, The process of generating a multi-channel spatial modulation map includes the following steps: The scene transition component, action coherence component, and causal necessity intensity component are subjected to time-domain alignment and smoothing filtering. The time domain alignment refers to mapping each component value to the time axis of video generation based on the time positions of the preceding event node and the subsequent event node in the structured event sequence, and ensuring that each component value decays to a preset baseline level during time intervals without event coverage. The smoothing filtering process is used to eliminate numerical abrupt changes in the modulation map in the time dimension caused by event node switching, and to generate a multi-channel spatial modulation map that evolves continuously in time.

6. The natural language-driven video generation method based on intent deconstruction according to claim 4, characterized in that, Before performing element-wise multiplication of the multi-channel spatial modulation map with the attention weight matrix, the following steps are also included: Based on the causal necessity intensity component in the logical influence factor vector, determine whether there is a key frame interval in the video segment corresponding to the subsequent event node where the causal effect should be significantly visualized. If so, within the keyframe interval, based on the magnitude of the causal necessity intensity component, the intensity of the spatial channel region associated with the subsequent event node and key object in the multi-channel spatial modulation map is increased to generate an enhanced spatial modulation map.

Images