Short video automatic generation method and system based on semantic comprehension

Abstract

The invention relates to a short video automatic generation method and system based on semantic understanding, and belongs to the technical field of intelligent content generation. The method comprises the following steps: acquiring multi-modal source data input by a user, and extracting a structured semantic feature vector; obtaining a dynamic semantic hypergraph through conflict detection and dynamic expansion of associated nodes; executing a Markov decision process on the dynamic semantic hypergraph to generate a platform instruction set; generating a video element sequence by applying a combination principle, presetting a conflict element pair, and outputting an enhanced video template; scheduling a layered material library according to the enhanced video template; and performing collaborative rendering on the layered material data flow through a heterogeneous computing architecture, and outputting a final generated video. And generating a quality feedback signal based on the behavior data of the user for the finally generated video, and dynamically updating the semantic deconstruction model and the association node of the dynamic semantic hypergraph. According to the invention, high automation and intelligence of short video content production are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent content generation, in particular to a short video automatic generation method and system based on semantic understanding. BACKGROUND

[0002] With the popularization of 5G networks and the improvement of mobile terminal performance, short videos have become one of the main carriers of information dissemination. The demand of users for content diversity, creative novelty and personalized expression continues to grow. The traditional generation method relying on manual editing or template splicing has been difficult to meet the production needs of massive customized content. Under this background, building an intelligent system that can understand user creation intent and automatically complete the whole process from semantic analysis to video output has become an important research direction in the field of multimedia artificial intelligence.

[0003] However, the existing short video automatic generation technology still faces many deep technical bottlenecks in actual application, which restricts its practicability and reliability in complex semantic environment. At present, the understanding of the existing video automatic generation system for input information often stays at the surface feature matching level, lacking the ability to deeply mine the implied emotional tendency and cultural symbol meaning. Especially in the scene of complex emotional expression or cultural background embedding, the existing system can only identify the surface keywords, but cannot deeply understand the emotional color, social cognition and collective memory symbols behind the language, resulting in semantic deviation between the generated content and the user's expectation. Moreover, due to the significant differences in aspect ratio, frame rate specification, code rate limitation and content style preference of different short video platforms, and the static template library matching strategy adopted by the existing technology cannot adaptively adjust according to the dynamic propagation law of the target platform, resulting in a decline in user experience.

[0004] Therefore, how to build a short video generation system that can deeply integrate multi-modal semantic understanding and has cross-platform adaptation ability is a technical problem to be solved at present. SUMMARY

[0005] In order to solve the above technical problems, the present application provides a short video automatic generation method and system based on semantic understanding.

[0006] In a first aspect, the present application provides a short video automatic generation method based on semantic understanding, which adopts the following technical solution: A short video automatic generation method based on semantic understanding, the short video automatic generation method comprising: acquiring multi-modal source data input by a user, extracting a structured semantic feature vector through a semantic deconstruction model; constructing a hypergraph and a hyperedge based on the structured semantic feature vector, obtaining a dynamic semantic hypergraph by conflict detection and dynamic expansion of associated nodes; According to the target platform identifier and the preset platform rule library, a Markov decision process is performed on the dynamic semantic hypergraph to generate a platform instruction set; Based on the dynamic semantic hypergraph and the platform instruction set, a combination principle is applied to generate a video element sequence, and a conflict element pair is preset, and an enhanced video template is output; According to the enhanced video template, a hierarchical material library is scheduled to generate a hierarchical material data stream; A final generated video is output by performing cooperative rendering on the hierarchical material data stream through a heterogeneous computing architecture; Based on the behavior data of the user for the final generated video, a quality feedback signal is generated to dynamically update the semantic deconstruction model and the associated nodes of the dynamic semantic hypergraph.

[0007] By adopting the above technical solution, a complete closed loop chain from multi-modal semantic understanding to intelligent video generation to user feedback driven optimization is constructed, semantic conflict detection and cultural adaptation are realized through a dynamic semantic hypergraph, platform customization parameter generation is completed with the help of a Markov decision process, creative structure design is guided by an enhanced template, and efficient cooperative rendering is realized relying on a GPU / NPU / FPGA heterogeneous architecture. Finally, the model performance is continuously optimized driven by user feedback, realizing high automation and intelligence of short video content production.

[0008] Optionally, the step of constructing a hypergraph and a hyperedge based on the structured semantic feature vector, detecting conflicts and dynamically expanding associated nodes to obtain a dynamic semantic hypergraph includes: The structured semantic feature vector is mapped to an initial graph node; A hyperedge is created to connect the initial graph nodes that have semantic association; Symbolic conflicts and logical contradictions between the initial graph nodes are detected to generate a conflict label set; According to the conflict label set, the knowledge graph is queried to replace the conflict nodes with compatible symbolic nodes; Based on the historical behavior data of the user, the associated nodes are dynamically expanded to generate a dynamic semantic hypergraph structure.

[0009] By adopting the above technical solution, the knowledge graph driven semantic reasoning and the user behavior guided dynamic evolution are combined to realize intelligent reconstruction and continuous optimization of the semantic graph structure. Multi-modal semantic aggregation is realized through cross-dimension hyperedges, semantic contradictions are accurately located using the conflict label set and the evidence chain mechanism, symbolic replacement is completed without destroying the topological relationship by using the graph structure equivalent replacement strategy, and personalized semantic expansion is realized based on historical behavior data. This technical solution not only solves the deficiencies of traditional methods in cultural adaptability, logical consistency and personalized expression, but also improves the depth and flexibility of semantic modeling.

[0010] Optionally, the step of generating the platform instruction set according to the target platform identification and the preset platform rule library, and performing a Markov decision process on the dynamic semantic hypergraph comprises: obtaining a target platform identification and a preset platform rule library; wherein the preset platform rule library comprises a frame ratio, a frame rate range, and a special effect density constraint; mapping the node topology structure of the dynamic semantic hypergraph to a state space of the Markov decision process; defining an action space according to the preset platform rule library; wherein the action space comprises a combination of frame parameters, frame rate parameters, and special effect parameters; evaluating the platform adaptation degree of the state-action pair through a reward function, and iteratively selecting an action with the maximum reward value; outputting a platform instruction set comprising the selected frame parameters, frame rate parameters, and special effect parameters.

[0011] By using the above technical solution, the topological information of the dynamic semantic hypergraph is used as the state input of the Markov decision process, breaking through the traditional coarse-grained decision mode based on content tags, and realizing fine-grained perception of the semantic association structure; through the double-factor reward function, the content coverage and platform matching degree are simultaneously optimized, solving the inherent contradiction between creative integrity and transmission adaptability; and with the help of reinforcement learning methods such as Q-learning, autonomous optimization of parameter combination is realized, significantly improving the intelligent level of decision-making.

[0012] Optionally, based on the dynamic semantic hypergraph and the platform instruction set, a combination principle is applied to generate a video element sequence, and a conflict element pair is preset, and the step of outputting an enhanced video template comprises: analyzing the node association relationship of the dynamic semantic hypergraph, and calculating the semantic association degree between nodes; applying the combination principle to merge node groups with a semantic association degree exceeding a preset association degree threshold, to generate an initial video element sequence; identifying a conflict element pair with an antagonistic relationship in the dynamic semantic hypergraph, to generate a conflict element identification set; presetting a time-space separation channel for each conflict element pair in the conflict element identification set; optimizing the rendering priority of the initial video element sequence according to the special effect parameters in the platform instruction set, to obtain a rendering priority identifier; outputting an enhanced video template comprising the initial video element sequence, the time-space separation channel of the conflict element pair, and the rendering priority identifier.

[0013] By adopting the technical scheme, the topological relationship of the dynamic semantic hypergraph is converted into operable visual arrangement instructions, time-optimal alignment of semantic aggregation is realized through a dynamic montage mechanism, multi-type semantic confrontation problems are solved by means of a three-dimensional conflict separation strategy, and a hierarchical rendering control model is established in combination with platform special effect parameters, thereby realizing intelligent transition from semantic intention to an executable template.

[0014] Optionally, the step of generating the hierarchical material data stream according to the enhanced video template includes: parsing the initial video element sequence and the rendering priority identifier in the enhanced video template; mapping a hierarchical loading strategy according to the rendering priority identifier; wherein the hierarchy includes a foreground layer, a background layer and a special effect layer; retrieving a hierarchical material library based on the initial video element sequence to generate a material identifier set of each level; performing preloading and format conversion of the hierarchical material according to the hardware constraint parameter of the platform instruction set; allocating independent storage buffers to conflict element pairs existing in the time-space separation channel; outputting the hierarchical material data stream containing the hierarchical association relationship.

[0015] By adopting the technical scheme, the three-dimensional dynamic priority mapping is used to realize differentiated scheduling of the foreground, the background and the special effect, the hardware constraint parameter is used to drive signal level conversion operations such as frame rate interpolation and anisotropic scaling, and the adaptive storage buffer mechanism is constructed based on the conflict type, thereby realizing efficient conversion from the "enhanced template" to the "executable data stream", achieving a high balance between resource utilization efficiency, content expression accuracy and cross-platform compatibility, and providing a solid and reliable resource scheduling foundation for automated video generation.

[0016] Optionally, the step of performing cooperative rendering on the hierarchical material data stream by a heterogeneous computing architecture to output a finally generated video includes: receiving the hierarchical material data stream and parsing the foreground layer dynamic model, the background layer scene data and the special effect layer particle parameter therein; allocating heterogeneous computing resources according to rendering task complexity; performing physical engine solving on the special effect layer particle parameter to generate particle motion trajectory and collision detection data; performing style transfer operation on the background layer scene data to adapt to the frame parameter of the platform instruction set; performing skeleton animation driving and light synthesis on the foreground layer dynamic model; coordinating rendering progress of each level through a timestamp synchronization mechanism; The synthesized picture is subjected to code rate compression by the FPGA encoder, and a final generated video meeting the platform requirements is output.

[0017] By adopting the technical scheme, optimal configuration of computing resources is achieved by using a task-hardware dynamic binding mechanism, authenticity of special effects and naturalness of subject actions are respectively ensured based on a physical engine and inverse kinematics, artistic expressiveness is improved by combining style transfer and light synthesis, and stability of multi-hardware collaboration is ensured by using timestamp synchronization and dynamic LOD degradation, so that high unification is achieved among physical authenticity, visual expressiveness and engineering efficiency, and high-performance and highly adaptive underlying rendering support is provided for cross-platform automated video generation.

[0018] Optionally, based on the behavior data of the user for the final generated video, a quality feedback signal is generated, and the steps of dynamically updating the semantic deconstruction model and the associated nodes of the dynamic semantic hypergraph include: Interaction behavior data of the user and the final generated video is collected, and a viewing time length, an interactive operation frequency and a content preference label are extracted; A quality feedback signal is calculated according to the behavior data, and an update weight coefficient of the semantic deconstruction model is generated; The parameter vector of the semantic deconstruction model is adjusted based on the update weight coefficient, and an optimized semantic feature extraction rule is output; The associated nodes of the dynamic semantic hypergraph are reconstructed according to the optimized semantic feature extraction rule, and a reconstructed node is obtained; Logical conflicts of the reconstructed node and an existing hyperedge structure are detected, and an updated dynamic semantic hypergraph is output after conflict resolution.

[0019] By adopting the technical scheme, a nonlinear quality feedback signal is generated by using multi-modal behavior data fusion, controlled parameter updates of the semantic deconstruction model are driven, and the optimized feature rules are applied to node reconstruction of the dynamic semantic hypergraph. In this process, a cultural compatibility verification mechanism based on a knowledge graph is introduced, realizing a closed-loop evolution of “learning-reconstruction-correction”. This technical scheme not only overcomes the problems of single feedback and easy semantic conflicts in traditional updating methods, but also balances model adaptability and cultural safety, ensuring that the generation system continuously evolves while always meeting user cognitive rules and social and cultural norms, providing a sustainable and reliable knowledge evolution basis for intelligent content generation.

[0020] Optionally, after the step of outputting the enhanced video template, the method further includes: Interest label sets in pre-stored user portrait data are read; Semantic similarities of each feature node in the dynamic semantic hypergraph and each label in the interest label set are respectively calculated, and a maximum similarity value is selected as an association degree index of the feature node. screening feature nodes with a correlation index exceeding a predetermined threshold, and constructing a narrative subgraph based on the connection relationship between the feature nodes and hyperedges; analyzing a video duration constraint parameter included in the platform instruction set to generate a set of branch paths in the narrative subgraph that satisfy a cumulative duration not exceeding the video duration constraint parameter; assigning a selection weight to each branch path and outputting a set of narrative paths with weighted identifiers.

[0021] By adopting the above technical solutions, the semantic vector space model is used to realize high-precision alignment of user interest and semantic nodes, the hyperedge structure is used to ensure the integrity of the narrative logic, and the feasible path set is generated under the dual boundaries of time duration constraint and cultural safety. Finally, intelligent optimization is realized through multi-factor weight distribution.

[0022] Optionally, the step of assigning a selection weight to each branch path includes: detecting path jump records in the user historical behavior log and increasing the weight value of the jumped path; when the branch path contains a high-priority element marked by the platform instruction set, the path base weight level is improved; for conflict element pairs with spatiotemporal separation channels, a conflict resolution factor is introduced in the weight calculation.

[0023] By adopting the above technical solutions, the historical jump records are detected to enhance the user preference response, the high-priority elements are identified to ensure the achievement of the operation target, and the conflict resolution factor is used to realize risk-sensitive creative inclusion, so that the system establishes a dynamic balance between personalization, functionality and security. The finally outputted set of narrative paths with weighted identifiers provides a solid content skeleton for subsequent enhanced video template instantiation, hierarchical rendering and cross-platform distribution, significantly improving the adaptability, compliance and user experience consistency of the short video automatic generation system.

[0024] In a second aspect, the present application provides a short video automatic generation system based on semantic understanding, which adopts the following technical solutions: A short video automatic generation system based on semantic understanding, the short video automatic generation system comprises: a data acquisition module configured to acquire multi-modal source data input by a user and extract structured semantic feature vectors through a semantic deconstruction model; a dynamic semantic hypergraph construction module configured to construct hypergraphs and hyperedges based on the structured semantic feature vectors, detect conflicts and expand associated nodes dynamically to obtain a dynamic semantic hypergraph; a platform instruction set generation module configured to execute a Markov decision process on the dynamic semantic hypergraph according to a target platform identifier and a preconfigured platform rule library to generate a platform instruction set; The video template enhancement module is used to generate a video element sequence based on the dynamic semantic hypergraph and the platform instruction set, apply the principle of combination, preset conflicting element pairs, and output an enhanced video template. The media scheduling module is used to schedule the layered media library according to the enhanced video template and generate a layered media data stream; The rendering module is used to perform collaborative rendering on the layered material data stream through a heterogeneous computing architecture and output the final generated video. The feedback update module is used to generate quality feedback signals based on user behavior data for the final generated video, and dynamically update the associated nodes of the semantic deconstruction model and the dynamic semantic hypergraph.

[0025] In summary, this application includes at least one of the following beneficial technical effects: by combining multimodal semantic understanding with hypergraph modeling, it realizes intelligent end-to-end processing from user input to automatic short video generation, can adaptively optimize the generation strategy according to the characteristics of the target platform, and continuously improves the generation quality and user satisfaction through layered rendering and feedback learning mechanisms, significantly improving the intelligence level and personalized adaptation capability of short video generation. Attached Figure Description

[0026] Figure 1 This is a first flowchart illustrating a semantic understanding-based method for automatically generating short videos, according to one embodiment of this application.

[0027] Figure 2 This is a second flowchart illustrating a semantic understanding-based method for automatically generating short videos, according to one embodiment of this application.

[0028] Figure 3 This is a schematic diagram of the third process of a semantic understanding-based automatic short video generation method according to one embodiment of this application.

[0029] Figure 4 This is a schematic diagram of the fourth process of a semantic understanding-based automatic short video generation method according to one embodiment of this application.

[0030] Figure 5 This is a schematic diagram of the fifth process of a semantic understanding-based automatic short video generation method according to one embodiment of this application.

[0031] Figure 6 This is a schematic diagram of the sixth process of a semantic understanding-based automatic short video generation method according to one embodiment of this application.

[0032] Figure 7 This is a schematic diagram of the seventh process of a short video automatic generation method based on semantic understanding, according to one embodiment of this application.

[0033] Figure 8This is a schematic diagram of the eighth process of a semantic understanding-based automatic short video generation method according to one embodiment of this application.

[0034] Figure 9 This is a ninth flowchart illustrating a semantic understanding-based method for automatically generating short videos, according to one embodiment of this application. Detailed Implementation

[0035] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-9 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0036] This application discloses a method for automatically generating short videos based on semantic understanding. The automatic generation method includes: Reference Figure 1 A method for automatically generating short videos based on semantic understanding, the method includes: Step S101: Obtain multimodal source data input by the user, and extract structured semantic feature vectors through a semantic deconstruction model; Among them, the structured semantic feature vector includes explicit semantic features, implicit emotional features, and cultural symbol features; Specifically, multimodal source data typically includes various heterogeneous data forms such as text descriptions, speech information, image or video clips, and user tags, which together constitute a complete expression of creative intent. In order to extract high-level semantic information that can be used for video generation from these complex inputs, this application introduces a semantic deconstruction model, a composite deep learning architecture. This model is not a single neural network, but a collaborative processing system that integrates multiple sub-modules.

[0037] Explicit semantic features refer to the content themes and keywords explicitly expressed by users, such as directly identifiable themes like "Spring Festival," "reunion dinner," and "fireworks." These are efficiently extracted using a quantized BERT model. Quantumization technology compresses the model parameter scale without significantly sacrificing accuracy, improving inference speed and making it suitable for real-time processing in mobile or edge computing scenarios. Implicit sentiment features focus on the emotional tendencies that users do not explicitly express but are implied in language style, tone changes, or visual emotions, such as joy, nostalgia, and anxiety. These features are highly subjective and context-dependent, so LSTM-ATT (Long Short-Term Memory Network combined with Attention Mechanism) is used for modeling. LSTM can capture long-term dependencies in time series, while the attention mechanism gives the model higher weights to key sentiment words (such as "moved" and "shocked"), thereby accurately quantifying the intensity and polarity of emotions.

[0038] Next, cultural symbol features involve deeper elements of social cognition and collective memory, such as dragons symbolizing power, red representing joy, and lanterns signifying a festive atmosphere. These symbols often transcend language boundaries and have fixed mapping relationships in different contexts. Therefore, association recognition is achieved through a pre-constructed knowledge graph. This knowledge graph stores a semantic network of "entity-relationship-entity" in the form of triples (such as <Spring Festival, usage, red envelope>) and supports reasoning extensions to achieve symbol matching in cross-cultural contexts.

[0039] Step S102: Construct a hypergraph and hyperedges based on structured semantic feature vectors, and obtain a dynamic semantic hypergraph by conflict detection and dynamic expansion of associated nodes; Among them, the nodes of the dynamic semantic hypergraph correspond to semantic features, and the hyperedges are associated with multi-dimensional feature nodes; For example, traditional graph structures only support binary relationships (i.e., an edge connects two nodes), making it difficult to express multi-dimensional semantic co-occurrence phenomena (such as "fireworks + music + countdown" jointly constituting a New Year's Eve scene). Hypergraphs, on the other hand, allow a single hyperedge to connect multiple nodes, naturally adapting to the needs of multimodal semantic aggregation. In the above framework, each semantic feature (such as "fireworks," "joyful," and "red") exists as an independent node in the graph, and hyperedges are dynamically generated based on semantic co-occurrence frequency, hierarchical relationships, or association paths in the knowledge graph. For example, the abstract concept of "festival celebration" can be connected to multiple specific nodes such as "fireworks," "drum music," and "crowd" through a single hyperedge, forming a semantic cluster.

[0040] More importantly, this hypergraph possesses a "dynamic" attribute, meaning that conflict detection and node expansion mechanisms are continuously executed during the construction process: Conflict detection aims to identify semantic contradictions or cultural taboos. For example, "white" symbolizes purity in Western weddings, but in the Chinese context, it is often used for funerals. If the input data contains both "Chinese wedding" and "white as the main color," the system will trigger a conflict determination and automatically query the knowledge graph to find compatible alternative symbols (such as replacing it with "ivory white" or "champagne"), thereby avoiding cultural misinterpretation; Dynamic expansion completes potential related nodes through knowledge graph reasoning. For example, after recognizing "Mid-Autumn Festival," it automatically adds derivative nodes such as "moon," "mooncake," and "longing," enhancing the breadth of semantic coverage.

[0041] Understandably, this dynamic evolution capability makes semantic hypergraphs not only static knowledge representation tools, but also cognitive engines with semantic reasoning and error correction capabilities, providing rich and consistent semantic context support for subsequent decision-making processes.

[0042] Step S103: Based on the target platform identifier and the pre-set platform rule base, perform a Markov decision process on the dynamic semantic hypergraph to generate a platform instruction set; The platform instruction set includes aspect ratio parameters, frame rate parameters, and special effects parameters.

[0043] Specifically, different short video platforms have significant differences in content presentation. For example, some platforms prefer vertical 9:16 aspect ratio, fast-paced editing and strong special effects, while some platforms, although they also support vertical screens, accept horizontal screen content and have higher requirements for frame rate and bit rate.

[0044] To this end, the system has a built-in platform rule base that stores the technical specifications (such as maximum file size, recommended encoding format, and frame rate range) and content preference models (such as average shot duration distribution and special effects usage density) of each platform.

[0045] Building upon this, Markov Decision Processes (MDPs) are used to model the sequential decision-making problem in video generation. The state space is defined as the joint representation of the current semantic hypergraph's node set and platform constraints, reflecting the semantic integrity and compliance the system possesses at a given moment. The action space contains a series of discretized video parameter adjustment operations, such as selecting a frame rate of 15fps or 30fps, setting the shot duration to 1 second or 3 seconds, and enabling or disabling motion blur effects. The design of the reward function is particularly crucial and can be implemented using a weighted combination approach. R = α × node coverage + β × platform matching degree; In the above formula, node coverage is used to measure how many semantic features are successfully mapped to video elements (avoiding information omission), platform matching degree is used to evaluate the degree of fit between the generated parameters and the best practices of the target platform (ensuring dissemination efficiency), and coefficients α and β can be adjusted according to user preferences (such as emphasizing creative integrity or dissemination adaptability). Through policy iteration or reinforcement learning methods such as Q-learning, the system gradually optimizes the decision path and finally outputs a set of optimal platform instructions, including specific parameters such as aspect ratio, frame rate configuration, transition frequency, and effects density, realizing the intelligent transformation from "semantic intent" to "platform compatibility".

[0046] Step S104: Based on the dynamic semantic hypergraph and platform instruction set, the video element sequence is generated by applying the combination principle, and conflicting element pairs are preset to output an enhanced video template. The combination principle refers to a system of mathematical rules for arranging video elements (such as shots, subtitles, sound effects, and animations) in an orderly manner based on semantic similarity, temporal logic, and visual harmony. An enhanced video template is a structured script with metadata annotations, containing semantic tags, durations, hierarchical affiliation (foreground / background / effects) for each shot, and interaction rules with other elements, providing precise navigation for subsequent material scheduling.

[0047] In this embodiment, the system first calculates the correlation between semantic features. If the semantic distance between "fireworks" and "cheers" is less than a preset distance threshold... If the elements are combined into a dynamic montage unit and presented synchronously on the timeline, then the elements that are antagonistic or visually disruptive (such as the combination of rapidly flashing lights and dense text which may cause visual fatigue) are treated as conflicting element pairs. The system preloads a spatiotemporal separation channel, which avoids cognitive overload by using time misalignment (showing the lights first and then the text) or spatial partitioning (placing the text in the safe area at the bottom of the screen).

[0048] In addition, the system can also introduce template creative entropy value as a quantitative indicator to measure the novelty of the video. Too high an entropy value indicates that the content is chaotic and disorderly, while too low an entropy value means that there is a lack of variation and surprise. The order and combination of elements can be adjusted by genetic algorithm or gradient optimization method so that the creative entropy of the final template converges to the ideal range.

[0049] Step S105: Based on the enhanced video template, schedule the layered media library to generate a layered media data stream; The layered material library is a multimedia resource pool organized according to visual hierarchy, including a foreground layer, a background layer, and a special effects layer. The foreground layer mainly stores the main content such as character motion models, dynamic text animations, and product display sequences. The background layer covers spatial supporting elements such as static images, environmental videos, and stylized background images. The special effects layer centrally manages components that enhance visual expressiveness, such as particle systems (e.g., sparks, raindrops, halos), filter presets, and transition animations.

[0050] It should be noted that the scheduling process should strictly follow the frame rate requirements in the platform instruction set and the element sequence in the enhanced template: for example, when the platform instruction set specifies 30fps, the system prioritizes loading the high frame rate version of the foreground dynamic model to ensure smooth action; when the semantic features contain "Chinese style" cultural symbols, the background layer automatically triggers the theme image generated based on GAN (such as ink wash landscape, classical window lattice patterns); and the number of particles in the effects layer is not a fixed configuration, but is dynamically adjusted according to the creative entropy value of the template—high entropy value scenes increase particle density to enhance visual impact, while low entropy value scenes reduce the use of effects to maintain a simple aesthetic.

[0051] This on-demand scheduling mechanism not only improves resource utilization efficiency but also ensures a balance between artistic style and technical performance in the generated content, forming a well-structured and hierarchical multi-layered data stream that lays the foundation for subsequent parallel rendering.

[0052] Step S106: Perform collaborative rendering on the layered material data stream through a heterogeneous computing architecture to output the final generated video; Among them, the GPU cluster processes particle effects, the NPU cluster performs style transfer, and the FPGA encoder outputs a video stream that meets the platform's bitrate requirements. Specifically, GPU clusters focus on processing computationally intensive graphics tasks, especially the parallel simulation and rendering of particle effects. They utilize thousands of CUDA cores to simultaneously calculate the position, velocity, color decay, and other attributes of millions of particles, achieving cinematic visual effects. NPU (Neural Processing Unit) clusters are responsible for performing style transfer operations on the background layer, such as converting ordinary street scene videos into Van Gogh painting style or cyberpunk color tones in real time. These tasks heavily rely on the inference capabilities of deep convolutional networks, and NPUs have advantages over CPUs / GPUs in terms of energy efficiency and latency. FPGA (Field Programmable Gate Array) encoders are responsible for the compression and encapsulation of the final generated video. Their reconfigurable logic circuits can dynamically adjust encoding parameters according to the bitrate specifications of different platforms, ensuring image quality while meeting transmission bandwidth limitations.

[0053] Step S107: Generate quality feedback signals based on user behavior data for the final generated video, and dynamically update the semantic deconstruction model and the associated nodes of the dynamic semantic hypergraph.

[0054] Among them, behavioral data includes, but is not limited to, completion rate, like rate, comment sentiment, forwarding path, and viewing device type. This data is aggregated into quality feedback signals.

[0055] In some embodiments, when the quality feedback signal is below a preset threshold, the system initiates a reverse diagnostic process: if the problem stems from semantic deconstruction bias (e.g., misjudging "nostalgia" as "sadness"), the BERT and LTM-ATT model parameters are fine-tuned through an online learning mechanism; if the problem lies in missing semantic associations (e.g., failing to recognize the strong association between "Mid-Autumn Festival" and "moon gazing"), the new relationship is written into the knowledge graph and the hypergraph construction rules are updated. This iterative optimization based on real user feedback enables the system not only to adapt to changes in individual preferences but also to capture the evolution of socio-cultural trends (e.g., emerging internet buzzwords and popular visual styles), thereby continuously improving the relevance and appeal of the generated content.

[0056] The above implementation constructs a complete closed-loop chain from multimodal semantic understanding to intelligent video generation and then to user feedback-driven optimization. It realizes semantic conflict detection and cultural adaptation through dynamic semantic hypergraph, completes platform-customized parameter generation with the help of Markov decision process, guides creative structure design with enhanced template, and realizes efficient collaborative rendering based on GPU / NPU / FPGA heterogeneous architecture. Finally, it continuously optimizes model performance with user feedback as the driving force, realizing a high degree of automation and intelligence in short video content production.

[0057] Reference Figure 2As one implementation of step S102, the steps of constructing a hypergraph and hyperedges based on structured semantic feature vectors, and obtaining a dynamic semantic hypergraph through conflict detection and dynamic expansion of associated nodes include: Step S201: Map the structured semantic feature vectors to initial graph nodes; In this system, each initial graph node corresponds to a single semantic feature. During the mapping process, each independent semantic unit in the vector is decomposed and instantiated as a node in the hypergraph, forming the initial graph structure. Each node carries its original semantic type label (explicit / sentiment / symbolic), weight coefficient, and source confidence. This fine-grained node-based processing not only preserves the integrity of the original semantics but also provides a structured foundation for subsequent multi-dimensional association analysis, enabling semantic elements of different modalities to interact and reason within a unified graph space.

[0058] Step S202: Create initial graph nodes with semantic associations that are connected by hyperedges; Among them, hyperedges contain cross-dimensional feature relationships. As a generalized edge structure, hyperedges allow an edge to connect multiple nodes at the same time, thus naturally modeling multi-dimensional semantic combination relationships.

[0059] In this embodiment, the system determines whether to establish a hyperedge by calculating the semantic correlation S between nodes, and the formula is a weighted average similarity: ; Here, `sim()` is the feature similarity function, employing different metrics for different feature types. For example, cosine similarity based on word embeddings (such as BERT vectors) is used for explicit semantic entities, Euclidean distance is used to map sentiment features to the [0,1] interval, and cultural symbols are evaluated based on path distance or semantic role matching in the knowledge graph. The weight coefficient `w`... i This reflects the importance of each dimension in the current context. For example, cultural symbols have a higher weight in celebratory content, while explicit semantic weight is dominant in news broadcasts.

[0060] When the calculated S exceeds a preset threshold θ, the system creates a hyperedge between the relevant nodes. This hyperedge not only records the set of connected nodes but also stores metadata about its internal multidimensional mapping relationships. For example, a triple association consisting of "Messi (dominant) — joy 0.92 (emotion) — ribbon rain (symbol)" represents the core semantic structure of a sports awards ceremony. This cross-dimensional hyperedge mechanism breaks through the limitations of traditional single-modal associations, achieving accurate modeling of complex semantic scenarios and providing rich contextual basis for subsequent conflict detection.

[0061] Step S203: Detect symbol conflicts and logical contradictions between nodes in the initial graph and generate a conflict marker set; Symbolic conflicts mainly stem from cultural context misalignment or violation of environmental constraints. For example, "rainbows" as a celebratory symbol have positive connotations in joyous occasions such as weddings and award ceremonies, but their appearance in funeral or mourning videos constitutes a strong cultural offense. The system identifies such conflicts by comparing the mutual exclusion rule base between cultural symbol nodes and their co-occurring environments (such as background music being funeral music and emotional polarity being sadness). Logical contradictions focus on the inconsistency between semantic actions and emotional states. For example, the explicit semantic is "running," a high-intensity action, but the emotional polarity is labeled as "fatigue -0.7." If the two co-occur without a reasonable explanation (such as "marathon runners who persevere to the end"), it is judged as a logical break.

[0062] In this embodiment, the detection process is not a simple rule matching, but rather the construction of a "mutually exclusive evidence chain," which involves extracting multi-hop reasoning paths from the knowledge graph that support conflict judgments, such as "funeral → somber atmosphere → no noise → ribbon rain symbolizing noise → conflict established." The final output conflict marker set is organized in the form of structured triples: {conflict type, conflict node ID, mutually exclusive evidence chain}, which clearly indicates the problem while retaining traceable reasoning basis, providing precise navigation for subsequent automated repair.

[0063] Step S204: Query the knowledge graph based on the conflict marker set and replace the conflicting nodes with compatible symbol nodes; When the system identifies a node (such as v_old="rain of ribbons") that needs to be removed due to cultural conflict, it does not simply delete the node, causing a semantic break. Instead, it initiates a search process in the knowledge graph's compatible symbol library to find alternative symbols that have similar functions in the same context but do not conflict. For example, "petals falling" or "lights gradually brightening" can both convey a solemn and warm atmosphere, suitable for funeral scenarios.

[0064] It's important to note that the replacement operation adheres to a strict principle of graph topology preservation: removing old nodes, adding new nodes, and inheriting all hyperedge connections from the original nodes to ensure the overall semantic structure of the graph doesn't collapse due to local modifications. For example, if the original hyperedge E connects the nodes "ribbon rain," "applause," and "victory," after the replacement, this hyperedge automatically redirects to "petals falling," maintaining the integrity of the "celebration-positive-visual feedback" semantic chain. This mechanism not only avoids the inefficiency and subjectivity of manual intervention but also automates and verifies semantic repair, significantly enhancing the cultural sensitivity and social acceptability of the generated content.

[0065] Step S205: Dynamically expand associated nodes based on the user's historical behavior data to generate a dynamic semantic hypergraph structure.

[0066] Among them, user historical behavior data covers their past interaction records such as watching, liking, completing, and sharing. The system extracts high-frequency co-occurrence semantic combination patterns through frequent itemset mining algorithms (such as Apriori or FP-Growth). For example, if a user has a long-term preference for the combination of "starry sky + soothing music + slow motion", then the "starry sky" node will be given a higher expansion priority.

[0067] For example, when the frequency freq(vk) of a certain symbol vk exceeds a preset threshold σ, the system adds it as an extended node v_k to the current hypergraph node set V, and recalculates its connection probability with existing nodes based on semantic relevance, establishing new hyperedges to enhance semantic richness. This expansion is not blind stacking, but guided by the user interest model, injecting personalized elements while maintaining the stability of the semantic backbone. The final output dynamic semantic hypergraph structure conforms to universal cultural norms and reflects individual aesthetic preferences, becoming a bridge connecting general semantic understanding and personalized content generation.

[0068] The above implementation combines knowledge graph-driven semantic reasoning with dynamic evolution guided by user behavior, achieving intelligent reconstruction and continuous optimization of the semantic graph structure. By achieving multimodal semantic aggregation through cross-dimensional hyperedges, accurately locating semantic contradictions using conflict marker sets and evidence chain mechanisms, and employing a graph structure equivalent replacement strategy to complete symbol replacement without disrupting topological relationships, this solution not only addresses the shortcomings of traditional methods in terms of cultural adaptability, logical consistency, and personalized expression, but also improves the depth and flexibility of semantic modeling.

[0069] Reference Figure 3 As one implementation of step S103, the step of performing a Markov decision process on the dynamic semantic hypergraph to generate a platform instruction set based on the target platform identifier and a pre-set platform rule base includes: Step S301: Obtain the target platform identifier and the preset platform rule library; wherein, the preset platform rule library includes aspect ratio, frame rate range and special effects density constraints; Among them, the target platform identifier (such as "Platform A - Portrait Mode", "Platform B - Landscape Compatibility" or "Platform C - Low Bitrate Adaptation") is not just a simple string label, but a key index that triggers the system to load the corresponding platform's technical specifications and content preferences.

[0070] The pre-built platform rule base is a structured collection of metadata that stores the hard requirements and recommended practices of each short video platform in terms of visual presentation and technical transmission: Aspect ratio constraints define the legal range of video aspect ratios. For example, Platform A mainly promotes the 9:16 vertical screen format to adapt to full-screen browsing on mobile devices, while Platform B supports vertical screens but also accepts 4:5 or even 1:1 ratios; Frame rate range specifies the acceptable minimum and maximum frame rates (such as 30fps to 60fps), which directly affects video smoothness and file size; Special effects density constraints limit the frequency of use of visual elements such as transitions, filters, and dynamic text per unit time to avoid playback stuttering or review restrictions due to over-rendering.

[0071] It should be noted that these rules do not exist in isolation, but are encoded into a computable set of constraints C using a formal syntax. p It covers multiple expression methods such as numerical range, Boolean conditions and weighted scoring, ensuring that the system can perform accurate technical alignment based on semantic understanding.

[0072] Step S302: Map the node topology of the dynamic semantic hypergraph to the state space of a Markov decision process; Specifically, in traditional video generation systems, states are typically simplified to content category tags (such as "food," "travel," and "dance"), lacking a detailed depiction of internal semantic relationships. In contrast, this application innovatively uses a dynamic semantic hypergraph G(V,E) as the core carrier of states, where V represents a set of nodes composed of explicit semantics, emotional polarity, and cultural symbols, and E is a set of hyperedges connecting multidimensional features, together forming a topological structure rich in semantic associations and contextual dependencies. This graph structure not only records "which semantics exist" but also expresses "how they are related"—for example, "fireworks" and "cheers" are connected by hyperedges, indicating a celebratory scene; the co-occurrence of "loneliness" and "night" suggests a contemplative atmosphere.

[0073] It is understandable that this complete topology information, along with platform constraints C, p Together they form a composite state space S={G(V,E),C} p This design endows each state with both semantic integrity and environmental adaptability. This gives the system "semantic structure awareness," meaning that when making parameter decisions, it can not only consider the current content theme but also judge the complexity and expressive needs of the content based on the connection density between semantic nodes, emotional consistency, and cultural compatibility. This allows for technical choices that better align with the original creative intent; for example, high-emotion-density scenes may require higher frame rates to enhance the coherence of emotional delivery.

[0074] Step S303: Define the motion space according to the preset platform rule library; wherein, the motion space includes a combination of frame parameters, frame rate parameters and special effects parameters; Specifically, the frame parameter set includes the scale a. ratio (e.g., 9:16, 4:5, 16:9) and direction a orientation (Portrait, Landscape, Adaptive) The frame rate parameter set covers the minimum frame rate b minFPS With maximum frame rate b maxFPS (e.g., 24fps, 30fps, 60fps), the special effects parameters are determined by density c. density (Low, Medium, High) and Type c type (Particles, transitions, and filters) together form a multidimensional discrete parameter combination space.

[0075] In this embodiment, the system generates a finite but comprehensive set of action candidates A={a1,a2,...,a...} by enumerating these parameters through Cartesian product and combining them with platform compliance filtering. n Each action 'a' represents a specific video technology configuration scheme. This rule-driven action space construction method avoids the computational waste caused by blind searching, ensures the engineering feasibility of the output instructions, and provides a clear evaluation dimension for the design of the subsequent reward function, so that each decision can find a balance between "creative expression" and "platform compatibility".

[0076] Step S304: Evaluate the platform adaptability of the state-action pair through the reward function, and iteratively select the action with the largest reward value; The reward function is defined as follows: ; In the above formula, the first term Coverage(V) a This measures the proportion of semantic nodes that are effectively expressed under the current action 'a', reflecting the degree of preservation of content information. For example, choosing a low frame rate may result in dynamic emotional features not being fully displayed, thereby reducing coverage. The second item Match(a,C) p This quantifies the degree of matching between the action and platform constraints, and achieves refined evaluation through a hierarchical verification mechanism: Image aspect ratio compliance d1 uses an indicator function to determine whether it is within the allowed range (e.g., 9:16 is 1 in the platform rules, and 16:9 is 0), the formula is:

[0077] The frame rate matching degree d2 is calculated as the overlap rate between the actual frame rate range and the platform-supported range. The formula is:

[0078] When there is complete overlap, d2=1; when there is no overlap, d2=0. The final matching degree is obtained by weighted combination Match(a,C). p The equation is derived from w1d1 + w2d2.

[0079] The entire reward function achieves dynamic shifts in the optimization focus by adjusting the coefficients a and β (satisfying α+β=1), thereby increasing a when emphasizing creative integrity and enhancing β when pursuing platform adaptability, demonstrating the system's strategic flexibility.

[0080] This reward signal is used to guide the iterative learning process of the Q-learning algorithm, and the update formula is: ; In the above formula, the learning rate η controls the speed at which new experiences are absorbed, and the discount factor γ measures the importance of future rewards. Through continuous trial and error and value accumulation, the system gradually converges to the optimal strategy and selects the action that can obtain the maximum long-term reward in the current state.

[0081] Step S305: Output a platform instruction set containing selected aspect ratio parameters, frame rate parameters, and effects parameters.

[0082] Specifically, this instruction set is not a single parameter, but a set of structured configuration commands that are directly input into the subsequent video rendering engine to control key aspects such as the frame cropping area, frame rate interpolation algorithm, and the intensity of special effects layer overlay.

[0083] For example, if the system ultimately selects the action as "9:16 aspect ratio, 60fps frame rate, high effect density", the instruction set will explicitly instruct the rendering module to enable super-frame technology to ensure the smoothness of high-speed action and activate the particle system to enhance the visual impact.

[0084] More importantly, this decision-making process is not statically preset, but dynamically generated based on the real-time state of the current semantic hypergraph. It has a high degree of context sensitivity and personalized adaptability, and can automatically adjust technical strategies according to different content themes, such as prioritizing frame rate for dance videos and focusing on aspect ratio information density for text and image videos.

[0085] In the above implementation, the topological information of the dynamic semantic hypergraph is used as the state input of the Markov decision process, which breaks through the traditional coarse-grained decision-making mode based on content tags and realizes fine-grained perception of semantic association structure; by simultaneously optimizing content coverage and platform matching degree through a two-factor reward function, the inherent contradiction between creative integrity and dissemination adaptability is resolved; and by using reinforcement learning methods such as Q-learning, the autonomous optimization of parameter combination is realized, which significantly improves the intelligence level of decision-making.

[0086] Reference Figure 4As one implementation of step S104, the steps of generating a video element sequence based on a dynamic semantic hypergraph and platform instruction set, applying the principle of combination, pre-setting conflicting element pairs, and outputting an enhanced video template include: Step S401: parse the node association relationships of the dynamic semantic hypergraph and calculate the semantic association degree between nodes; In this context, nodes represent explicit semantic entities (such as "fireworks"), emotional polarities (such as "joy +0.9"), or cultural symbols (such as "Spring Festival red envelopes"), while hyperedges connect multidimensional feature combinations with co-occurrence, causal, or metaphorical relationships.

[0087] Based on this, the system constructs a weighted similarity model R. ab To accurately measure the semantic affinity between any two nodes a and b, the formula is as follows: ; In the above formula, The weight coefficients represent the k-th dimension features (such as text semantics, sentiment intensity, and symbol category), reflecting the importance of that dimension in the current context; sim() is a similarity function designed for different types of features. For explicit semantics, cosine similarity based on BERT embedding is used; for sentiment features, normalized Euclidean distance is used; and for cultural symbols, path semantic similarity in the knowledge graph (such as based on WordNet or a dedicated cultural ontology) is used for measurement; while I k As a node interaction frequency factor, co-occurrence statistics from user historical behavior data are introduced. For example, if "fireworks" and "countdown" frequently appear synchronously in past successful videos, then their I... k An increase in the value further strengthens the correlation score between the two.

[0088] Understandably, this multi-dimensional weighted fusion mechanism not only considers static semantic matching but also incorporates dynamic usage patterns, making the correlation calculation both consistent with linguistic logic and close to actual dissemination patterns, thus providing a scientific basis for subsequent element combinations.

[0089] Step S402: Apply the principle of combination to merge node groups whose semantic relevance exceeds a preset relevance threshold to generate an initial video element sequence; Among them, when the semantic correlation degree R between multiple nodes ab When the preset relevance threshold t1 is reached, it indicates that these elements constitute a semantically close aggregation unit. For example, "drum music," "lion dance," and "red decorations" all point to the theme of "traditional festivals." The system then triggers a combination mechanism to encapsulate them into a composite sequence of video elements. This process is not a simple stacking but rather involves performing timeline alignment optimization. , In the above formula, This represents the optimal presentation time for each node i under ideal conditions (e.g., "fireworks" should erupt at the climax, and "blessings" should appear at the end). The optimal unified playback time T is obtained by minimizing the sum of squared errors between the ideal time and the actual alignment time of each element. align This achieves coordination and unity of multiple elements across the time dimension.

[0090] The aforementioned mathematical modeling method ensures that even if the original semantic elements differ in rhythm, an initial video sequence with a smooth rhythm and progressive emotion can be formed through intelligent arrangement, avoiding rhythmic breaks or misalignment of focus caused by traditional template-based splicing. The generated initial video element sequence already possesses a basic narrative structure, but potential conflicts still need further processing to ensure the logical consistency of the content expression.

[0091] Step S403: Identify conflicting element pairs with adversarial relationships in the dynamic semantic hypergraph and generate a set of conflicting element identifiers; The antagonistic relationships include not only explicit logical mutual exclusion, but also implicit emotional conflicts and cultural taboos. The system can detect these by constructing a composite criterion function. ; In the above formula, the first term I (Polarity) a ≠Polarity b Determine whether the emotional polarities of the two nodes are opposite. For example, the co-occurrence of "cheerful music" and "sad text" constitutes an emotional conflict; the second item Query the predefined set of compatible symbols (CompatSet) in the knowledge graph to determine whether a cultural symbol is allowed to coexist with another symbol. For example, "white wedding dress" is usually not paired with "black decorations" in a Chinese wedding because their symbolic meanings are contradictory.

[0092] Once a conflict is identified, the system generates a structured identifier set, recording the conflict type label (such as "emotional conflict", "spatial conflict", "cultural symbol conflict"), the IDs of the involved nodes, and mutually exclusive chains of evidence, forming a traceable and manageable conflict list.

[0093] Step S404: Preset a spatiotemporal separation channel for each conflicting element pair in the conflicting element identifier set; For example, for pairs of elements with conflicting emotions (such as background music with a cheerful melody but subtitles expressing sorrow), the system allocates independent audio tracks and shared video tracks, allowing sound and image to run separately in terms of emotional expression. Users can switch between mute and video to perceive different emotional levels, avoiding cognitive confusion caused by the contrast between sound and image. For spatial conflicts (such as "indoor tea room" and "torrential rain" being unable to coexist in the same physical scene), a picture-in-picture layer segmentation technique is used to embed the "rain scene" as a floating window in the corner of the main screen, preserving the imagery connection while avoiding spatial paradoxes. For conflicts of cultural symbols (such as the sensitive juxtaposition of "cross" and "Buddha statue" in a specific religious context), the system loads a temporal and spatial interlaced presentation sequence, that is, alternately displaying the two symbols with transition animations or black screen intervals inserted in between, using the temporal misalignment to achieve semantic decoupling.

[0094] Understandably, this multimodal separation mechanism not only maintains the existence of the original semantic elements, but also resolves their antagonism through structured arrangement, significantly enhancing the inclusiveness and artistic tension of content expression.

[0095] Step S405: Optimize the rendering priority of the initial video element sequence according to the special effects parameters in the platform instruction set to obtain the rendering priority identifier; Among them, the special effects density parameter c in the platform instruction set density (Such as "Low", "Medium", "High") are no longer just global switches, but are rendering levels that participate in the dynamic calculation of particle effects: ; In the above formula, N node The formula, representing the number of active semantic nodes in the current video segment, ensures that the allocation of special effects resources is proportional to the content complexity—automatically increasing particle levels to enhance visual impact in high-density semantic scenes (such as celebratory climaxes) and reducing special effects usage in concise narrative segments to ensure smooth performance. Furthermore, rendering priority is also influenced by the spatiotemporal separation channel: isolated conflicting elements receive independent layer weights, ensuring their separation logic is faithfully reproduced during rendering. The entire optimization process achieves a dynamic balance between content importance, platform capabilities, and user experience, enabling the generated template to both meet technical constraints and maximize expressive effects.

[0096] Step S406: Output an enhanced video template containing the initial video element sequence, spatiotemporal separation channels of conflicting element pairs, and rendering priority identifiers.

[0097] This enhanced video template not only defines "what to play" but also clarifies "how to play," including the temporal order of elements, spatial layout, conflict resolution paths, and rendering weights at each level, providing a complete execution guide for subsequent material scheduling and collaborative rendering.

[0098] Understandably, this "enhanced" feature is reflected in three dimensions: semantic enhancement (improving the coherence of expression through montage combination), logical enhancement (ensuring semantic consistency through spatiotemporal separation), and technical enhancement (achieving precise resource allocation through priority optimization), forming a highly intelligent, scalable intermediate representation for video generation with self-correction capabilities.

[0099] In the above implementation, the topological relationships of the dynamic semantic hypergraph are transformed into operable visual orchestration instructions. The dynamic montage mechanism achieves optimal temporal alignment of semantic aggregation. A three-dimensional conflict separation strategy is used to solve multi-type semantic adversarial problems. A hierarchical rendering control model is established in combination with platform special effects parameters, realizing the intelligent leap from semantic intent to executable template.

[0100] Reference Figure 5 The steps for generating a layered media data stream based on an enhanced video template scheduling layered media library include: Step S501: Parse the initial video element sequence and rendering priority identifier in the enhanced video template; The system first performs syntactic parsing on the template, identifies the visual level (foreground, background, or effect) of each element, and extracts its associated priority parameters. For example, a dynamic character model is marked as high priority and time-sensitive, while a certain ink painting style background emphasizes the weight of cultural symbols.

[0101] Step S502: Map the layered loading strategy according to the rendering priority identifier; wherein, the layers include the foreground layer, the background layer, and the effects layer; For the foreground layer, which typically carries the main content such as character actions, product displays, or dynamic text, its loading priority is determined by a combination of rendering flags (indicating whether real-time compositing is required) and time sensitivity (time critical, such as the requirement for keyframes to appear on time), ensuring that the core narrative elements are fully ready before playback. The background layer is responsible for creating the atmosphere, and its priority is entirely determined by the importance of the cultural symbols it carries. For example, the "red walls of the Forbidden City" have extremely high weight in a traditional Chinese style theme and must be loaded first to ensure stylistic consistency. The effects layer, due to its computational intensity and conflict relevance, requires a combination of particle density (reflecting the complexity of the effects) and conflict indicator I. conflict The conflict indicator factor is set to 1 when a spatiotemporal separation channel exists, triggering a higher priority and ensuring that the special effects resources of conflicting elements can be prepared in time to support the implementation of the separation logic.

[0102] Step S503: Based on the initial video element sequence, retrieve the hierarchical material library and generate material identifier sets for each level; Specifically, the system accesses the corresponding dedicated material library according to the established hierarchical classification. The foreground layer queries the dynamic model library, which stores 3D / 2D driveable models with skeletal animation, facial expression parameters, and motion capture data. The search results return the model's unique identifier (such as UUID) and the corresponding animation parameters (such as blendshape weights and motion trajectory curves). Next, the background layer connects to the environment material library, which contains high-resolution static images, panoramic video clips, and lighting presets (HDR environment maps), and outputs scene identifiers and lighting configuration parameters for subsequent lighting consistency restoration in the rendering engine; the effects layer matches the particle system library, which integrates various preset emitters (such as flames, snowflakes, and light dust) and their physics engine parameters (gravity coefficient, lifecycle, and collision response).

[0103] Step S504: Based on the hardware constraint parameters of the platform instruction set, perform preloading and format conversion of layered materials; Among them, the frame rate parameter b in the platform instruction set maxFPS This directly affects the playback smoothness of the foreground dynamic model. If the original model only supports 24fps while the target platform requires 60fps, the system will automatically perform frame interpolation calculations. ; That is, inserting a frame between every two original frames. insert An intermediate frame, generated by optical flow or a deep network, is used to perform temporal signal resampling, ensuring smooth motion at high frame rates. Similarly, the aspect ratio parameter `aratio` triggers a non-uniform scaling operation on the background scene: ; By adjusting the horizontal scaling ratio, the original aspect ratio is adapted to the target ratio (e.g., from 16:9 to 9:16 portrait mode), avoiding content cropping or black borders. These conversion operations are all completed during the preloading phase, utilizing idle computing cycles to process resources in advance, significantly reducing the pressure on real-time rendering.

[0104] More importantly, all conversions follow the principle of "fidelity first," preserving the original semantic features as much as possible while changing the format. For example, content-aware fill technology is used during scaling to protect the main area from distortion.

[0105] Step S505: Allocate independent storage buffers for conflicting element pairs that have spatiotemporal separation channels; For elements of emotional conflict (such as the coexistence of cheerful music and sad subtitles), the system allocates a dual-channel audio buffer, allowing the left and right channels to carry audio tracks with different emotions, and users can perceive the emotional tension by switching between stereo. Elements of spatial conflict (such as the main screen being an indoor tea room and the sub-screen being an outdoor rainstorm) are allocated to a picture-in-picture layer storage area, with the main background and sub-background stored in independent texture buffers, which are then mixed and output by the synthesizer according to layer weights. Elements of cultural symbol conflict (such as alternating religious totems) are loaded with a sequence of frames with staggered timestamps, with each frame containing only one symbol, and the alternating presentation of "one waxing and the other waning" is achieved through precise time scheduling to avoid cultural offense caused by simultaneous display.

[0106] Step S506: Output a layered material data stream containing hierarchical relationships.

[0107] This layered material data stream not only contains the original material data (models, textures, particle parameters) for each layer, but also encapsulates the spatial stacking order, temporal synchronization relationships, conflict handling paths, and buffer mapping information between layers, forming a standardized input that can be directly consumed by heterogeneous rendering architectures. The preservation of these "relationships" allows subsequent computing units such as GPUs and NPUs to work collaboratively in a unified coordinate system, avoiding layer misalignment or logical breaks caused by missing information.

[0108] In the above implementation, three-dimensional dynamic priority mapping is used to achieve differentiated scheduling of foreground, background and special effects. Based on hardware constraint parameters, frame rate interpolation and non-isotropic scaling signal-level conversion operations are driven. An adaptive storage buffer mechanism is built based on conflict type. This achieves efficient conversion from "enhanced template" to "executable data stream", thereby achieving a high balance between resource utilization efficiency, content expression accuracy and cross-platform compatibility, and providing a solid and reliable resource scheduling foundation for automated video generation.

[0109] Reference Figure 6 As one implementation of step S106, the step of performing collaborative rendering on the layered material data stream through a heterogeneous computing architecture to output the final generated video includes: Step S601: Receive the layered material data stream and parse the foreground layer dynamic model, background layer scene data, and special effects layer particle parameters. The system first performs protocol parsing on the data stream, identifying the data format, resource identifiers, and metadata tags at each level: the foreground layer contains dynamic model identifiers, skeletal topology, and animation parameters, used to drive the movement of characters or objects; the background layer encapsulates scene images, lighting configurations, and style transfer targets, determining the overall visual tone; and the effects layer provides particle system parameters, such as emission rate, lifespan, and initial velocity field, used to generate dynamic visual elements such as flames, rain, snow, and light effects.

[0110] This parsing process not only extracts the raw data, but also reconstructs the spatial stacking relationship and time synchronization requirements between different layers, providing a unified task view for the subsequent parallel processing of heterogeneous hardware, ensuring that different computing units can operate efficiently in their respective domains while maintaining the consistency of the global rendering target.

[0111] Step S602: Allocate heterogeneous computing resources according to the complexity of the rendering task; Among them, the GPU cluster is bound to the special effects layer rendering task, the NPU cluster is bound to the background layer rendering task, and the FPGA encoder is bound to the video stream output task. Specifically, this application's embodiments abandon the traditional "one-size-fits-all" approach to unified GPU rendering and instead construct a task-hardware mapping table to accurately match the computational characteristics of the task with the hardware's strengths.

[0112] For example, when the number of detected particles exceeds the threshold N p When there are 1000 particles, the system automatically assigns the special effects layer task to the GPU cluster for processing. This is because the GPU cluster has powerful parallel floating-point computing capabilities, making it suitable for performing physical simulations of large-scale particle systems. If the scene complexity C... s If the value exceeds a preset threshold (e.g., 5.0) through a comprehensive evaluation of indicators such as mesh density, texture resolution, and number of illumination layers, the NPU cluster will be activated to perform style transfer operations, using its tensor computation units optimized for neural networks to accelerate deep learning model inference; while the video encoding task is always undertaken by the FPGA encoder, because its programmable logic structure supports low latency and high throughput real-time bitrate control.

[0113] Understandably, this dynamic binding mechanism not only improves resource utilization but also prevents high-load tasks from blocking the critical path, achieving optimal allocation of computing power. More importantly, this mapping strategy has condition-triggered adaptability, enabling it to dynamically adjust task allocation based on real-time load, such as automatically downgrading to CPU-based software rendering on low-end devices, ensuring cross-platform compatibility.

[0114] Step S603: Perform physics engine calculations on the particle parameters of the special effects layer to generate particle motion trajectories and collision detection data; The system simulates the physical interaction behavior between particles by solving a constrained optimization problem. ; In the above formula, Δ xi m represents the position correction amount for the i-th particle. i For its quality, C j (x) represents the constraint function (such as distance constraint, collision constraint, incompressibility constraint) defined on the particle system. λij These are the corresponding Lagrange multipliers, used to gradually satisfy physical constraints during the iteration process.

[0115] Compared to traditional force-based dynamics simulations (such as the Newton-Euler method), this technical solution offers higher numerical stability and real-time performance, making it particularly suitable for frame rate-sensitive scenarios in video generation. After the calculation is complete, the system outputs the particle position matrix P. out =[x1, ,x k This process records the spatial coordinates of all particles in each frame, which are then used by the GPU shader for instantiation and rendering. This not only ensures the natural motion trajectories of special effects such as fireworks explosions and water splashes, but also prevents particles from penetrating the model or background through precise collision detection, significantly improving visual credibility.

[0116] Step S604: Perform a style transfer operation on the background layer scene data to adapt to the frame parameters of the platform instruction set; Style transfer is not simply a matter of layering filters, but rather involves loading a pre-trained deep neural network model G. style :R H×W×3 →R H×W×3 This model transforms original scene images into visual representations with specific artistic styles (such as ink painting, oil painting, or cyberpunk). It typically uses a convolutional neural network (CNN) or Transformer architecture to separate content and style representations in the feature space and achieves style transfer by optimizing the loss function. Furthermore, the system needs to adjust the aspect ratio 'a' according to the platform instruction set. ratio Perform adaptation transformation: ; The process involves first padding the stylized image to expand the canvas, then cropping it to the target aspect ratio. This ensures the final background retains artistic integrity while meeting platform display requirements. For example, when converting a 16:9 oil painting-style cityscape to a 9:16 portrait orientation, the system adds gradient transition areas at the top and bottom before cropping to prevent the main content from being truncated. This operation is efficiently executed by the NPU cluster, fully leveraging its advantages in matrix operations and memory bandwidth.

[0117] Step S605: Perform skeletal animation driving and lighting compositing on the dynamic model of the foreground layer; The system first loads the skeletal topology B=(J,E) based on the dynamic model identifier, where J is the set of joints and E is the bone connection edge.

[0118] Subsequently, regarding the target pose T target (e.g., the "waving" action), solve for the optimal joint angles. : ; Among them, F K (θ) is the positive kinematic function, describing the position of the end effector (such as the hand) at a given joint angle. This optimization problem can be solved efficiently using Jacobian matrix iteration or the Cyclic Coordinate Descent (CCD) algorithm, ensuring natural and smooth character movements.

[0119] Meanwhile, the system places the foreground model in a unified lighting environment and performs PBR (Physically Based Rendering) material calculations in conjunction with the HDR lighting parameters of the background layer, so that the shadows, highlights and reflections of the foreground objects are consistent with the background lighting, avoiding the unsightly feeling of "floating on the screen".

[0120] Step S606: Coordinate the rendering progress of each layer through a timestamp synchronization mechanism; Because heterogeneous units such as GPUs, NPUs, and FPGAs operate in different clock domains and have varying processing latencies, rendering delays at certain layers may cause screen tearing or audio-visual desynchronization. To address this, the system assigns a unified synchronous clock source to each rendering unit and monitors its task completion timestamps in real time. Once a delay Δt is detected at a certain level (e.g., a special effects layer experiencing excessive particle delays), the system immediately initiates a dynamic degradation strategy. ; This means reducing the level of detail at a given latency level, for example, by reducing the number of particles, lowering texture resolution, or simplifying shadow calculations, to minimize visual quality loss while maintaining a stable overall frame rate. This coordinated degradation mechanism enables flexible collaboration across hardware, allowing the system to maintain smooth output even under high load.

[0121] In step S607, the FPGA encoder performs bitrate compression on the synthesized image and outputs the final generated video that meets the platform requirements.

[0122] Specifically, the FPGA encoder is based on the target code rate R in the platform instruction set. target Constructing a dynamic bitrate control model: ; In the above formula, QP is a quantization parameter that controls the compressive strength, and R... actual This represents the actual bitrate of the current frame. The model adjusts the QP value based on the error exponent to achieve rapid convergence to the target bitrate.

[0123] More importantly, the system dynamically adjusts the QP based on frame complexity (such as motion amplitude and texture richness), using a high compression ratio for static images and retaining more details for highly dynamic scenes, thereby maximizing visual quality under bandwidth-constrained conditions. The final output video stream not only meets the platform's frame rate, aspect ratio, and bitrate requirements, but also maintains the complete visual effect of collaborative rendering at all levels.

[0124] In the above implementation, the optimal allocation of computing resources is achieved by using a task-hardware dynamic binding mechanism. The realism of special effects and the naturalness of the main body's movements are ensured by physical engine calculation and inverse kinematics, respectively. The artistic expressiveness is enhanced by combining style transfer and lighting compositing. At the same time, the stability of multi-hardware collaboration is ensured by timestamp synchronization and dynamic LOD degradation. Thus, a high degree of unity is achieved between physical realism, visual expressiveness and engineering efficiency, providing high-performance and highly adaptable underlying rendering support for cross-platform automated video generation.

[0125] Reference Figure 7 As one implementation of step S107, the step of generating a quality feedback signal based on user behavior data for the final generated video and dynamically updating the associated nodes of the semantic deconstruction model and the dynamic semantic hypergraph includes: Step S701: Collect user interaction data with the final generated video, and extract viewing duration, interaction frequency, and content preference tags; The system uses event tracking technology to collect multi-dimensional user behavior logs on the player in real time, including but not limited to the video completion rate (Rc), which is the ratio of actual viewing time to total viewing time. This is used to measure content attractiveness and narrative coherence; a high Rc value indicates that users are willing to consume the entire content, suggesting a reasonable semantic structure; element interaction density. This reflects the degree of user engagement with specific visual elements (such as bullet screen triggers, clickable icons, and dynamic subtitles). High interaction density usually corresponds to emotional climaxes or key information areas; and the emotional preference tag set is as follows: By clustering and summarizing user likes, comment sentiment analysis, and sharing behaviors, the system extracts the preferred emotional types (such as "nostalgia," "inspiration," and "relaxation"). These behavioral indicators not only capture users' explicit actions but also reveal their potential aesthetic tendencies and cognitive preferences through statistical modeling, forming the original basis for subsequent model updates.

[0126] Step S702: Calculate the quality feedback signal based on the behavioral data and generate the updated weight coefficients of the semantic deconstruction model; Specifically, a nonlinear evaluation function for multimodal fusion is established: ; In the above formula, L gTo generate the original set of sentiment tags carried by the video. First item. Using the hyperbolic tangent function to perform a saturation transformation on the viewing integrity rate, its S-curve characteristics make it possible for R... c The gain tends to level off as it approaches 1, effectively suppressing the risk of overfitting from high completion rates, while maintaining a sensitive response in the low to medium range; the second term log(1+D) i Logarithmic augmentation of interaction density amplifies significant changes in low-frequency interactions (such as from 0 to 1 click) while preventing weight explosion caused by high-frequency interactions, thus achieving dynamic range compression; the third term I(L p ∩L g The function is an indicator function that determines whether there is an intersection between the user's actual preference tags and the preset tags of the generated content. If there is, it outputs 1; otherwise, it outputs 0. This function is used to detect the deviation between "expected emotion" and "actual feeling".

[0127] In this embodiment, by adjusting the configuration of factors α, β, γ, and δ, the system can flexibly adjust the contribution weights of each dimension. For example, it can emphasize interactive feedback during the content testing phase and highlight viewing completion rate during the optimization of mature content. The final generated W update It is not a simple weighted average, but a composite signal that comprehensively reflects content quality, user engagement, and emotional matching, which directly determines the strength and direction of the semantic deconstruction model parameter update.

[0128] Step S703: Adjust the parameter vector of the semantic deconstruction model based on the updated weight coefficients, and output the optimized semantic feature extraction rules; The semantic deconstruction model is typically a deep neural network architecture (such as Transformer or graph neural network), and its parameter vector θ determines how semantic features are extracted from the original multimedia data. This application's embodiment employs a gradient incremental update strategy: ; In the above formula, η is the learning rate, and J(θ) is the semantic feature loss function (such as contrastive loss or triplet loss), which measures the distance between the current feature representation and the ideal semantic space. The key innovation lies in the introduction of W. update As a modulation factor for gradient updates, the model is made to make significant adjustments only when the user feedback is positive, while remaining stable when the feedback is ambiguous or negative, thus avoiding model drift caused by noisy data.

[0129] Understandably, the above mechanism enables feedback-driven adaptive learning, allowing the model to continuously absorb real user preferences and gradually optimize its understanding of cultural context, emotional expression, and visual symbols. For example, if multiple users exhibit high R-values ​​in videos related to "traditional festivals,"... c With D iFurthermore, if the preference tags include "reunion" and "red", the model will enhance the feature weights of elements such as "lantern" and "couplets", thereby improving the recognition accuracy of similar content in the future.

[0130] Step S704: Reconstruct the associated nodes of the dynamic semantic hypergraph according to the optimized semantic feature extraction rules to obtain the reconstructed nodes; Among them, the system traverses the existing dynamic semantic hypergraph and the feature weights are below the threshold. Weakly connected nodes (such as those that have not been activated for a long time or have semantic ambiguity, such as "blurred background" or "transition animation") are determined based on the new parameter θ. new Re-perform feature extraction: ; In the above formula, f extract The feature extraction function may include convolutional coding, attention aggregation, or multimodal fusion operations. The reconstructed nodes not only possess more accurate semantic embedding vectors but may also trigger new hyperedge connections. For example, the originally isolated nodes "dumplings" and "Spring Festival Gala" might be identified as strongly related by the new model because they both appear in high-feedback videos, thus forming a "Spring Festival culture" hyperedge. This process realizes a dynamic growth and aging mechanism for the semantic network, ensuring that the hypergraph structure remains synchronized with user cognitive trends and preventing knowledge stagnation.

[0131] Step S705: Detect logical conflicts between the reconstructed nodes and the existing hyperedge structure, and output the updated dynamic semantic hypergraph after conflict resolution.

[0132] Adding or reconstructing nodes may introduce logical contradictions, especially symbolic ambiguity in cross-cultural contexts. The system detects these contradictions by constructing a conflict criterion function. ; This involves determining whether the cultural symbol carried by a new node is outside the compatible set of its superedge e. For example, if a new node named "white flower" is identified and attempts to join the "wedding celebration" superedge, but this symbol symbolizes mourning in a specific culture, a conflict is triggered. At this point, the system activates the knowledge graph compatibility query mechanism to retrieve semantic equivalents from the cultural symbol knowledge base that satisfy both conditions: ; In the above formula, sim() measures semantic similarity (such as based on word vectors or image embeddings), I culture The system identifies the symbol's compliance within the target culture. It prioritizes semantically similar and culturally compatible alternatives, such as replacing "white flower" with "red rose" or "peony," thus preserving visual functionality while avoiding cultural offense. This knowledge-guided conflict resolution mechanism ensures semantic updates are both learning-capable and maintain safe boundaries.

[0133] In the above implementation, multimodal behavioral data fusion is used to generate a nonlinear quality feedback signal, which drives the controlled parameter update of the semantic deconstruction model. The optimized feature rules are then applied to the node reconstruction of the dynamic semantic hypergraph. A knowledge graph-based cultural compatibility verification mechanism is introduced during this process, achieving a closed-loop evolution of "learning-reconstruction-correction." This technical solution not only overcomes the problems of single feedback and easy semantic conflict in traditional update methods, but also achieves a balance between model adaptability and cultural security. It ensures that the generation system continuously evolves while always conforming to user cognitive patterns and socio-cultural norms, providing a sustainable and reliable knowledge evolution foundation for intelligent content generation.

[0134] Reference Figure 8 As a further implementation of the short video automatic generation method, after the step of outputting the enhanced video template, it also includes: Step S801: Read the set of interest tags from the pre-stored user profile data; The process begins by reading the set of interest tags from the pre-stored user profile data. These tags serve as the source of personalized information and are derived from the in-depth mining results of the user's historical behavior data analysis module. They include, but are not limited to, long-term viewing preferences (such as "traditional Chinese culture" and "technology and digital products"), short-term hot topics (such as "Spring Festival customs" and "World Cup events"), emotional tendencies (such as "nostalgia" and "inspiration"), and social interaction characteristics (such as "frequent sharing of parent-child content").

[0135] Understandably, these tags are not simply keyword stuffing, but rather structured semantic expressions constructed through natural language processing, cluster analysis, and knowledge graph mapping, possessing clear contextual boundaries and weight levels. The system retrieves the set of interest tags L for the current target user by accessing the user profile database. user ={l1,l2,...,l n This provides individual cognitive model support for subsequent content adaptation.

[0136] Step S802: Calculate the semantic similarity between each feature node in the dynamic semantic hypergraph and each label in the interest label set, and select the maximum similarity value as the correlation index of the feature node. In the dynamic semantic hypergraph, the feature nodes represent specific semantic entities, such as "lion dance performance", "electronic music", "red lantern", etc. Each node carries a multi-dimensional feature vector (text description, visual embedding, emotional polarity, cultural affiliation).

[0137] In this embodiment, the system employs a pre-trained semantic vector space model (such as a BERT-based, CLIP-based, or domain-specific fine-tuned model) to uniformly map node feature vectors and user interest tag vectors to the same high-dimensional semantic space, within which cosine similarity calculations are performed. This operation can capture deep semantic connections, such as the implicit link between "lion dance" and "traditional culture," even if the two do not directly overlap in their literal expressions. For each node, the system iterates through all interest tags, calculates their similarity, and takes the maximum value as the relevance index S for that node. node =max l∈Luser sim(v node ,v l This mechanism quantifies the attractiveness of a node to the current user. It avoids the coarse-grained judgment of "label matching equals relevance" in traditional recommendation systems, achieving fine-grained measurement based on semantic space continuity.

[0138] Step S803: Filter feature nodes whose correlation index exceeds a predetermined threshold, and construct a narrative subgraph based on the hyperedge connection relationship between feature nodes; Among them, the system uses S node > (like Using 0.6 as the criterion, highly relevant nodes are extracted from the original dynamic semantic hypergraph to form a candidate node set. Subsequently, based on the hyperedge connections of these nodes in the original hypergraph, a topological reconstruction is performed, inheriting their cultural symbol associations, temporal series dependencies, and emotional progression logic, to construct a locally connected narrative subgraph.

[0139] For example, if the three nodes "lion dance," "gongs and drums," and "temple fair" all meet the relevance threshold, and the original graph contains an emotionally progressive hyperedge of "lion dance → gongs and drums → temple fair," then this path is retained as part of the subgraph. In particular, this process strictly inherits the hyperedge structure across nodes, ensuring that the generated subgraph not only contains semantically related elements but also maintains the consistency of the logical chains and cultural context in the original semantic network, preventing narrative breaks or semantic distortion caused by isolated node extraction.

[0140] Step S804: Parse the video duration constraint parameters contained in the platform instruction set and generate a set of branch paths in the narrative subgraph that satisfy the condition that the cumulative duration does not exceed the video duration constraint parameters; Among them, the video duration constraint T in the platform instruction set max As a hard boundary condition, it participates in the path feasibility determination. The system traverses all possible node sequence combinations (i.e., potential narrative paths) in the narrative subgraph, and calculates the cumulative estimated duration for each path P=(n1,n2,...,nk). , where t iThe standard presentation duration for the i-th node (derived from template definition or historical statistics). Only when T... P ≤T max At that time, the path is included in the branch path set.

[0141] In addition, the system excludes paths containing nodes with conflicting cultural symbols, such as combinations that contain both "cross" and "Buddha statue" without a spatiotemporal separation mechanism. By querying a pre-built cultural compatibility knowledge base to perform logical filtering, the system ensures that each candidate path is semantically consistent and culturally compliant.

[0142] Step S805: Assign selection weights to each branch path and output a set of narrative paths with weight labels.

[0143] The system detects path jump records in users' historical behavior logs, identifies narrative patterns that users frequently choose or linger on in past videos (such as "first show the product → then demonstrate the usage scenario → finally present user reviews"), and adds weight values ​​to the current path that matches such patterns to reflect the influence of behavioral inertia.

[0144] Next, when a branch path contains elements marked as high priority by the platform instruction set (such as brand logos, promotional information, and policy publicity content), the system increases the basic weight level of that path to ensure that business or operational goals are met first.

[0145] Finally, for conflicting element pairs with spatiotemporal separation channels in the path (such as "upbeat music + sad text" presented through stereo separation), the system introduces a conflict resolution factor α∈(0,1], adjusting the weights according to the effectiveness of the separation strategy. If the separation mechanism has been verified to effectively alleviate cognitive conflict, the weight is appropriately increased; otherwise, it is decreased. The resulting set of narrative paths with weighted labels not only reflects the comprehensive interplay between user preferences and platform requirements but also provides sortable and schedulable decision inputs for subsequent dynamic rendering and A / B testing.

[0146] In the above implementation, a high-precision alignment between user interests and semantic nodes is achieved based on the semantic vector space model, the integrity of the narrative logic is ensured by using hyperedge structure inheritance, and a set of feasible paths is generated under the dual boundaries of time constraints and cultural security. Finally, intelligent optimization is achieved through multi-factor weight allocation.

[0147] Reference Figure 9 As a further implementation of the short video automatic generation method, the step of assigning selection weights to each branch path includes: Step S901: Detect path jump records in the user's historical behavior log and increase the weight value of the jump path; Among them, path jump records refer to non-linear browsing behaviors that users actively trigger when watching short videos in the past, such as jumping to a specific plot by clicking the progress bar, switching story branches by using interactive buttons, or repeatedly watching a certain segment in a multi-segment video.

[0148] This behavioral data is structured and stored in user behavior logs, and associated with specific narrative structure patterns. By analyzing these records, the system identifies narrative path patterns that users tend to prioritize or repeatedly experience, such as "suspenseful opening → plot twist → emotional climax" or "product demonstration → user testimonial." When a currently generated branch path is highly similar to these high-frequency jump paths in terms of topological structure or semantic sequence, the system incrementally increases its weight value.

[0149] For example, if a user repeatedly navigates to fitness video clips containing "before and after" comparisons, the system will increase the weight of sequences with "before and after weight loss comparison images + data annotations" in the current candidate path. This weight enhancement mechanism based on behavioral memory is essentially a form of implicit preference mining based on reinforcement learning, enabling the system to go beyond the surface association of label matching and deeply capture the user's narrative rhythm preferences and information acquisition habits.

[0150] Step S902: When a branch path contains a platform instruction set marked as a high-priority element, increase the basic weight level of the path. Platform instruction sets, acting as external control signals, are typically driven by needs such as operational strategies, advertising placement, policy compliance, or brand collaborations. These instructions explicitly specify that certain content elements must receive priority display, such as "brand logos must appear in the first 3 seconds," "specific promotional information must be presented in full," or "public service announcements must not be edited." These high-priority elements have a mandatory presence requirement within the narrative path, but their inclusion and placement still require overall evaluation.

[0151] The innovation of this application's embodiments lies in the fact that it does not use a hard insertion method to disrupt the narrative flow, but instead introduces it as a weight modulation factor into the selection mechanism: once a branch path fully contains one or more high-priority elements and its presentation conforms to the specifications (such as sufficient duration and visual clarity), the basic weight level of the path is systematically improved.

[0152] For example, if two paths are similar in terms of user interest matching, but one of them includes a platform-specified "new product launch countdown" animation, the latter will receive a higher overall weight, thus having a higher probability of being selected in the subsequent template filling and rendering stages. This soft guidance mechanism not only ensures the realization of the platform's core needs but also avoids blatant interference with the user experience, achieving an organic integration of business logic and content quality.

[0153] Step S903: For conflicting element pairs with spatiotemporal separation channels, a conflict resolution factor is introduced into the weight calculation.

[0154] Conflicting element pairs refer to combinations of nodes that have potential contradictions at the cultural, emotional, or logical levels, such as "red festive background" versus "black and white mourning attire," "upbeat music" versus "sad text," and "religious symbols" versus "entertainment scenes." If such elements are presented directly side-by-side, it may cause cognitive dissonance or cultural offense.

[0155] However, in some creative expressions, such conflicts themselves have artistic tension or rhetorical significance (such as irony, contrast, suspense), so they should not be simply eliminated, but should be resolved in a compliant manner through "temporal and spatial separation channels", such as allocating conflicting elements to different time segments, picture-in-picture areas, or dual-channel audio channels.

[0156] In this embodiment, for conflicting element pairs with spatiotemporal separation channels, the system introduces a conflict resolution factor α∈(0,1] when calculating the selection weight of the branch path to which it belongs. The value of α depends on the historical effectiveness verification of the separation mechanism: if a certain separation strategy (such as time interleaving) has been well-received by users in the past (manifested as high completion rates and low negative feedback), then α is close to 1, indicating that the path can still retain a high weight; conversely, if similar processing has caused users to abandon or complain, then α is significantly reduced, inhibiting the optimal probability of the path. Mathematically, this factor can be applied to the multiplicative adjustment term of the total path weight: , where k is the number of conflict pairs that have been successfully resolved in the path. This mechanism enables the system to encourage creative expression while maintaining sensitive control over semantic risks, forming a controllable path selection strategy.

[0157] In the above implementation, detecting historical navigation records enhances user preference responsiveness, identifying high-priority elements ensures the achievement of operational goals, and conflict resolution factors enable risk-sensitive creative inclusion, thus establishing a dynamic balance between personalization, functionality, and security. The final output set of weighted narrative paths provides a solid content framework for subsequent enhanced video template instantiation, layered rendering, and cross-platform distribution, significantly improving the adaptability, compliance, and user experience consistency of the short video automatic generation system.

[0158] This application also discloses a short video automatic generation system based on semantic understanding.

[0159] A short video automatic generation system based on semantic understanding, the short video automatic generation system includes: The data acquisition module is used to acquire multimodal source data input by the user and extract structured semantic feature vectors through a semantic deconstruction model. The dynamic semantic hypergraph construction module is used to construct hypergraphs and hyperedges based on structured semantic feature vectors. By detecting conflicts and dynamically expanding associated nodes, a dynamic semantic hypergraph is obtained. The platform instruction set generation module is used to perform a Markov decision process on the dynamic semantic hypergraph based on the target platform identifier and the pre-set platform rule base to generate the platform instruction set; The video template enhancement module is used to generate video element sequences based on dynamic semantic hypergraphs and platform instruction sets, applying the principle of combination, and pre-setting conflicting element pairs to output enhanced video templates. The media scheduling module is used to schedule the layered media library based on the enhanced video templates and generate layered media data streams; The rendering module is used to perform collaborative rendering of layered material data streams through a heterogeneous computing architecture, and output the final generated video. The feedback update module is used to generate quality feedback signals based on user behavior data in response to the final generated video, and dynamically update the associated nodes of the semantic deconstruction model and the dynamic semantic hypergraph.

[0160] The semantic understanding-based short video automatic generation system of this application embodiment can implement any of the above-mentioned short video automatic generation methods, and the specific working process of each module in the short video automatic generation system can refer to the corresponding process in the above-mentioned method embodiments.

[0161] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0162] In this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0163] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce a good effect.

[0164] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A short video automatic generation method based on semantic understanding, characterized in that, The short video automatic generation method comprises: obtaining multi-modal source data input by a user, extracting a structured semantic feature vector through a semantic deconstruction model; constructing a hypergraph and a hyperedge based on the structured semantic feature vector, detecting conflicts and dynamically expanding associated nodes to obtain a dynamic semantic hypergraph; performing a Markov decision process on the dynamic semantic hypergraph according to a target platform identifier and a preset platform rule library to generate a platform instruction set; based on the dynamic semantic hypergraph and the platform instruction set, applying the principle of combination to generate a video element sequence and presetting a pair of conflict elements to output an enhanced video template; scheduling a hierarchical material library according to the enhanced video template to generate a hierarchical material data stream; performing cooperative rendering on the hierarchical material data stream through a heterogeneous computing architecture to output a final generated video; based on behavior data of the user for the final generated video, generating a quality feedback signal and dynamically updating the semantic deconstruction model and the associated nodes of the dynamic semantic hypergraph. 2.The short video automatic generation method based on semantic understanding of claim 1, wherein, The step of constructing a hypergraph and a hyperedge based on the structured semantic feature vector, detecting conflicts and dynamically expanding associated nodes to obtain a dynamic semantic hypergraph comprises: mapping the structured semantic feature vector into initial graph nodes; creating a hyperedge to connect initial graph nodes that have semantic associations; detecting symbol conflicts and logical contradictions between initial graph nodes to generate a conflict marker set; querying a knowledge graph according to the conflict marker set to replace conflict nodes with compatible symbol nodes; dynamically expanding associated nodes based on historical behavior data of the user to generate a dynamic semantic hypergraph structure. 3.The short video automatic generation method based on semantic understanding of claim 1, characterized in that, The step of performing a Markov decision process on the dynamic semantic hypergraph according to a target platform identifier and a preset platform rule library to generate a platform instruction set comprises: obtaining a target platform identifier and a preset platform rule library; wherein the preset platform rule library includes frame ratio, frame rate range and special effect density constraints; mapping the node topology structure of the dynamic semantic hypergraph into the state space of the Markov decision process; defining an action space according to the preset platform rule library; wherein the action space includes combinations of frame parameters, frame rate parameters and special effect parameters; evaluating the platform adaptation degree of the state-action pair through a reward function, and iteratively selecting the action with the maximum reward value; outputting a platform instruction set containing selected frame parameters, frame rate parameters and special effect parameters.

4. The short video automatic generation method based on semantic understanding according to claim 3, characterized in that, The step of applying the principle of combination to generate a video element sequence and presetting a pair of conflict elements based on the dynamic semantic hypergraph and the platform instruction set to output an enhanced video template comprises: analyzing the node association relationship of the dynamic semantic hypergraph to calculate the semantic association degree between nodes; applying the principle of combination to merge node groups with a semantic association degree exceeding a preset association degree threshold to generate an initial video element sequence; identifying a pair of conflict elements with an antagonistic relationship in the dynamic semantic hypergraph to generate a conflict element identifier set; presetting a space-time separation channel for each pair of conflict elements in the conflict element identifier set; optimizing the rendering priority of the initial video element sequence according to the special effect parameters in the platform instruction set to obtain a rendering priority identifier; outputting an enhanced video template containing the initial video element sequence, the space-time separation channel of the pair of conflict elements and the rendering priority identifier. 5.The short video automatic generation method based on semantic understanding of claim 4, characterized in that, According to the enhanced video template scheduling layered material library, the step of generating a layered material data stream comprises: Parsing the initial video element sequence and the rendering priority identifier in the enhanced video template; Mapping a layered loading strategy according to the rendering priority identifier; wherein the layering comprises a foreground layer, a background layer and a special effect layer; Retrieving the layered material library based on the initial video element sequence to generate a set of material identifiers of each level; Performing preloading and format conversion of the layered material according to the hardware constraint parameters of the platform instruction set; Allocating independent storage buffers to conflicting element pairs with spatiotemporal separation channels; Outputting a layered material data stream containing hierarchical association relationships. 6.The short video automatic generation method based on semantic understanding of claim 5, wherein, Through the heterogeneous computing architecture, the step of performing collaborative rendering on the layered material data stream to output the final generated video comprises: Receiving the layered material data stream, parsing the foreground layer dynamic model, the background layer scene data and the special effect layer particle parameters therein; Allocating heterogeneous computing resources according to the rendering task complexity; Performing physical engine calculation on the special effect layer particle parameters to generate particle motion trajectories and collision detection data; Performing style transfer operation on the background layer scene data to adapt to the frame parameters of the platform instruction set; Performing skeleton animation driving and lighting synthesis on the foreground layer dynamic model; Coordinating the rendering progress of each level through a timestamp synchronization mechanism; Performing rate compression on the synthesized picture by an FPGA encoder to output the final generated video that meets the platform requirements.

7. The short video automatic generation method based on semantic understanding according to any one of claims 1 to 6, characterized in that, Based on the behavior data of the user for the final generated video, the step of generating a quality feedback signal and dynamically updating the associated nodes of the semantic deconstruction model and the dynamic semantic hypergraph comprises: Collecting the interaction behavior data of the user with the final generated video, extracting the viewing time, the interaction operation frequency and the content preference label; Calculating the quality feedback signal according to the behavior data to generate an update weight coefficient of the semantic deconstruction model; Adjusting the parameter vector of the semantic deconstruction model based on the update weight coefficient to output the optimized semantic feature extraction rule; Reconstructing the associated nodes of the dynamic semantic hypergraph according to the optimized semantic feature extraction rule to obtain reconstructed nodes; Detecting the logical conflict of the reconstructed nodes with the existing hyperedge structure, and outputting the updated dynamic semantic hypergraph after conflict resolution. 8.The method of claim 1, wherein, After the step of outputting the enhanced video template, it further comprises: Reading the interest label set in the pre-stored user portrait data; Calculating the semantic similarity between each feature node in the dynamic semantic hypergraph and each label in the interest label set, and selecting the maximum similarity value as the association degree index of the feature node; Screening feature nodes with an association degree index exceeding a predetermined threshold, and constructing a narrative subgraph based on the hyperedge connection relationship between the feature nodes; Parsing the video duration constraint parameters contained in the platform instruction set to generate a branch path set in the narrative subgraph that satisfies the cumulative duration not exceeding the video duration constraint parameters; Assigning a selection weight to each branch path to output a set of narrative paths with weight identifiers. 9.The method of claim 8, wherein, The step of assigning a selection weight to each branch path comprises: Detecting the path jump records in the user historical behavior log to increase the weight value of the jumped path; When the branch path contains a platform instruction set marked as a high-priority element, the path base weight level is promoted; For conflict element pairs with spatial and temporal separation channels, a conflict resolution factor is introduced in the weight calculation. 10.A short video automatic generation system based on semantic understanding, characterized in that, The short video automatic generation system comprises: A data acquisition module configured to acquire multi-modal source data input by a user, and extract a structured semantic feature vector through a semantic deconstruction model; A dynamic semantic hypergraph construction module configured to construct a hypergraph and a hyperedge based on the structured semantic feature vector, detect and dynamically expand associated nodes through conflict detection, and obtain a dynamic semantic hypergraph; A platform instruction set generation module configured to execute a Markov decision process on the dynamic semantic hypergraph according to a target platform identifier and a preset platform rule library, and generate a platform instruction set; A video template enhancement module configured to generate a video element sequence based on the dynamic semantic hypergraph and the platform instruction set, preset conflict element pairs, and output an enhanced video template; A material scheduling module configured to schedule a hierarchical material library according to the enhanced video template, and generate a hierarchical material data stream; A rendering module configured to perform cooperative rendering on the hierarchical material data stream through a heterogeneous computing architecture, and output a final generated video; A feedback updating module configured to generate a quality feedback signal based on behavior data of a user with respect to the final generated video, and dynamically update the semantic deconstruction model and associated nodes of the dynamic semantic hypergraph.

Images