Text-driven AIGC role action sequence automatic generation method

Abstract

The application relates to the technical field of action generation, and discloses an AIGC role action sequence automatic generation method based on text driving, which establishes an AIGC role action sequence automatic generation system. The system obtains original texts, action reference data, a physical parameter library and performance samples through parallel convergence of multi-source heterogeneous data. Fuzzy natural language is converted into structured action feature parameters by using six-dimensional semantic disintegration and double-branch decoupling mapping. Temporal-spatial composite arrangement optimization is carried out based on multi-branch parallel attention, time sequence priority and spatial constraint loss analysis, so that continuous 3D skeleton joint data is generated. After motion smoothing, foot correction, stylized enhancement and semantic-physical-biomechanics triple verification, the final product is packaged and exported in a standard format. The application realizes end-to-end automatic generation from abstract text to high-quality role action sequence, and solves the problems of semantic understanding deviation, rough time sequence arrangement, lack of spatial constraints and insufficient stage cooperation in the prior art.

Description

Technical Field

[0001] This invention relates to the field of motion generation technology, specifically to a text-driven method for automatically generating AIGC character motion sequences. Background Technology

[0002] With the rapid development of the digital content industry, the demand for automated generation of character motion sequences is becoming increasingly urgent in fields such as film and television production, game development, and virtual reality. Traditional motion generation methods mainly rely on motion capture equipment to collect data from live actors or on animators manually creating motion frame by frame. These methods suffer from significant problems such as high equipment costs, long production cycles, and difficulty in quickly responding to diverse textual description requirements.

[0003] In recent years, text-driven action generation technology has gradually emerged, but existing methods still have the following shortcomings: First, natural language text is highly abstract and ambiguous, and existing systems struggle to accurately understand the multi-dimensional semantic information contained in the text, such as action categories, attribute modifications, emotional states, and narrative logic, leading to significant semantic discrepancies between the generated actions and the text descriptions. Second, when processing the temporal structure of action sequences, existing methods lack refined modeling and priority scheduling mechanisms for long- and short-term dependencies between frames, making it difficult to achieve efficient sequence compression and rhythm arrangement while ensuring accurate reproduction of key actions. Third, existing methods generally neglect key factors such as the geometric rigidity constraints of the human skeleton, the avoidance of physical collisions of limbs, and biomechanical limit constraints during the action space generation process, resulting in unnatural phenomena such as joint stretching deformation, limb penetration, and joint over-limits in the generated actions. Fourth, existing systems operate independently of each module and lack collaborative optimization mechanisms, failing to form a complete integrated technical chain from text input to action sequence output.

[0004] Therefore, there is an urgent need for an AIGC character action sequence automatic generation method that can deeply integrate multi-source heterogeneous data, accurately quantify and analyze natural language semantics, integrate and coordinate spatiotemporal constraints and physical laws, and achieve end-to-end automated closed-loop generation. This method aims to break through the bottlenecks of existing technologies in semantic understanding, temporal arrangement, spatial constraints and system coordination, and meet the urgent needs of film and television production, game development, virtual reality and other fields for high-quality, high-efficiency and high-controllability character action generation. Summary of the Invention

[0005] 1. Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a text-driven method for automatically generating AIGC character action sequences. This method boasts advantages such as deep fusion of multi-source heterogeneous data, accurate quantitative translation of natural language semantics, integrated arrangement of spatiotemporal constraints, coordinated guarantee of physical authenticity and narrative consistency, and automated closed-loop delivery throughout the entire process. It solves the problems of semantic understanding bias, rough temporal arrangement, lack of spatial constraints, and insufficient stage coordination in existing technologies.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution: a text-driven method for automatically generating AIGC character action sequences, comprising the following steps:

[0009] Step 1: System Generation: Establish an AIGC character action sequence automatic generation system. The system includes a source text multi-dimensional feature acquisition module, an action semantic standardization translation module, a composite spatiotemporal arrangement analysis module, an action decoding generation module, an action sequence post-processing and verification module, and an action sequence generation output module.

[0010] Step 2, Data Acquisition: The source text multi-dimensional feature acquisition module aggregates raw text, motion reference data, physical parameter library, and performance samples from multiple sources;

[0011] Step 3, Semantic Translation: The action semantic standardization translation module receives multi-source aggregated data, performs standardized translation, and outputs action feature parameters;

[0012] Step 4, Composite Spatiotemporal Choreography: The composite spatiotemporal choreography analysis module performs composite choreography optimization under spatiotemporal constraints on the standardized translated motion feature parameters and outputs the optimal motion sequence parameters;

[0013] Step 5, Motion Decoding and Generation: The motion decoding and generation module receives the optimal motion sequence parameters output by the composite spatiotemporal orchestration analysis module, and maps the feature parameters into continuous 3D skeletal joint rotation and position data through the motion decoder to generate a preliminary motion sequence.

[0014] Step 6, Post-processing and Verification: The motion sequence post-processing and verification module performs motion smoothing filtering, foot landing constraint correction and stylistic detail enhancement on the preliminary motion sequence, and completes text-action semantic consistency verification, physical rationality verification and biomechanical safety verification.

[0015] Step 7, Final Output: The action sequence generation output module will encapsulate and export the verified final action sequence in a standard action data format, completing the entire process of automatic generation of AIGC character action sequences based on text-driven methods.

[0016] Preferably, the source text multidimensional feature acquisition module has a built-in natural language text acquisition unit, action reference data acquisition unit, physical parameter library acquisition unit and performance sample data collection unit; the composite spatiotemporal choreography analysis module has a built-in multi-branch parallel attention analysis unit, temporal priority analysis unit and spatial feature constraint loss analysis unit.

[0017] Preferably, the natural language text acquisition unit acquires natural language text through a user input interface / API text access method; the motion reference data acquisition unit acquires motion reference data through motion capture equipment and a public motion database.

[0018] Preferably, the physical parameter library acquisition unit acquires the physical parameter library through physical engine simulation sampling and human biomechanics literature calibration; the performance sample data collection unit collects performance sample data through film and animation datasets and professional actor motion capture special dynamics.

[0019] Preferably, the action semantic standardization translation module translates the feature extraction data using a six-dimensional semantic decomposition and dual-branch decoupling mapping method, specifically including the following process:

[0020] S1.1 Based on the preset fine-grained action lexicon and annotation rules, the input original text of the source text multi-dimensional feature acquisition module is decomposed into six fine-grained features: action subject, core posture, motion trajectory, speed level, force level and motion range. The fuzzy natural language description is then structurally completed and standardized through semantic completion and contextual reasoning technology based on a large language model.

[0021] S1.2 Then, a fine-grained semantic encoder is used to extract specific features from the decomposed multi-dimensional semantic features, and relying on a self-built text-action parameter mapping knowledge base, the text description is automatically mapped into accurately quantified basic action parameters.

[0022] S1.3 Finally, a semantic dual-branch decoupling architecture is adopted to separate action features into basic action ontology features and action attribute regulation features:

[0023] (1) Basic branch recognition of core movement types such as waving, raising hands, and walking:

[0024] (2) The attribute branches independently parse the semantics of the modifiers "forceful", "gentle", "fast" and "slow";

[0025] S1.4 Finally, output the quantitative control parameters of force, speed, amplitude, and duration, and complete the standardized translation of natural language text into structured and decoupled motion feature parameters.

[0026] Preferably, the multi-branch parallel attention analysis unit calculates global multimodal context feature weights based on different modal features of text semantics, action posture, and temporal position, combined with an adaptive weighted fusion algorithm. The calculation formula is as follows:

[0027]

[0028] In the formula, m represents the number of parallel attention branches. These represent the query matrix, This represents the query for the i-th branch. This represents the attention calculation function for the i-th branch. This represents the adaptive fusion weight of the i-th branch;

[0029] The multi-branch parallel attention analysis unit calculates the global multimodal context feature weights. The attention outputs of each branch are adaptively weighted and fused to generate a joint multimodal feature tensor that integrates global context semantics, pose structure and temporal position information.

[0030] Preferably, the temporal priority analysis unit calculates the long and short-term temporal dependencies between action frames and the inter-frame priority ranking based on the long and short-term temporal dependencies between different frames in the action sequence, combined with a priority constraint ranking algorithm. The calculation formula is as follows:

[0031]

[0032] In the formula, Indicates the first The final timing priority score of frame actions. , These represent the short-term independent contribution coefficient and the long-term dependent contribution coefficient, respectively. This represents the Sigmoid activation function. , Let represent the learnable weight matrix and bias term of the short-term feature linear mapping, respectively. This represents the action feature vector of frame t. This indicates the length of the context window for long-range dependency modeling. Represents the time-series decay factor. This represents the cosine similarity function, used to evaluate the features of the current frame. Features of historical frames The degree of similarity in their movements, postures, and trends.

[0033] Preferably, the timing priority analysis unit is based on the first Final timing priority score for frame actions Value distribution dynamically adjusts frame sampling density and interpolation strategy:

[0034] When the Final timing priority score for frame actions When ∈(0.7, 1.0]: mark the frame as a core keyframe, preserve its motion pose data completely, and do not perform downsampling;

[0035] When the Final timing priority score for frame actions When ∈ [0.3, 0.7]: Adaptive sampling and linear interpolation are performed on the frame, based on... The specific values ​​are retained proportionally in a portion of the frames;

[0036] When the Final timing priority score for frame actions When ∈[0, 0.3): the frame is determined to be a temporally redundant frame, and sparsification is performed. The transition posture is generated by Bézier curve / spline interpolation fitting using the preceding and following high / medium priority frames.

[0037] Preferably, the spatial feature constraint loss analysis unit calculates the spatial constraint loss value of the human skeleton's geometric structure, physical collision, and biomechanical limits based on the topological rigidity constraint of the human skeleton, the spatial penetration potential energy field, and the physiological limit function of joint angular momentum. The calculation formula is as follows:

[0038]

[0039] The calculation formulas for each sub-item are as follows:

[0040]

[0041]

[0042]

[0043] In the formula, This represents the total loss value due to spatial constraints. , , These represent the balance weighting coefficients for geometric stiffness loss, collision penetration loss, and angular momentum limit loss, respectively. The geometric stiffness loss term is used to constrain the bone length to be constant, where... For the skeleton topological edge set, This represents the total number of edges of the skeleton. , Let be the three-dimensional spatial coordinates of the a-th and b-th joints, respectively. This indicates the standard physiological length of the bone. Represents the spatial penetration loss term, where Where ϵ is the total number of joints, and ϵ is the set minimum safe distance threshold. , Penalties are only imposed on penetrations that are less than the safe distance. The term representing the physiological limit of joint angular momentum loss is given by the expression, where The total number of movable joints, This represents the current rotation angle of the k-th joint. , These represent the maximum and minimum physiologically permissible angles of the joint as specified in biomechanical literature.

[0044] Preferably, the spatial feature constraint loss analysis unit is based on the total spatial constraint loss value. The generated joint coordinates and rotation parameters are updated by gradient optimization through backpropagation algorithm, and iterated round by round until the loss converges to below the preset threshold, and finally outputs a spatially reasonable sequence of motion postures.

[0045] Compared with existing technologies, this invention provides a text-driven method for automatically generating AIGC character action sequences, which has the following advantages:

[0046] 1. This invention utilizes a multi-source heterogeneous data parallel aggregation mechanism through a source text multi-dimensional feature acquisition module to integrate four heterogeneous information sources: natural language text, motion capture data, physical simulation parameters, and performance sample data. It also relies on a unified feature representation framework to achieve cross-modal semantic alignment, thereby breaking through the bottleneck of one-sided information and style homogenization in motion generation caused by traditional single data source-driven approaches. This provides a complete data foundation covering semantic intent, motion laws, physical constraints, and artistic expression for subsequent semantic understanding and motion generation.

[0047] 2. This invention, through the six-dimensional semantic decomposition and dual-branch decoupling mapping mechanism of the action semantic standardization translation module, accurately converts fuzzy natural language descriptions into structured parameter vectors with six dimensions: action subject, core posture, motion trajectory, speed level, force level, and motion range. Furthermore, it employs an architecture design where the basic action subject branch and the action attribute control branch are independently parsed. This allows the method to overcome the core deficiency of existing technologies, which can only map text to coarse-grained action categories but cannot quantify and parse semantic modifications such as force, gentleness, speed, and slowness. Ultimately, it achieves the goal of standardizing and automatically adapting natural language text to precisely quantified and flexibly controllable action feature parameters.

[0048] 3. This invention overcomes the problems of disconnect between temporal arrangement and spatial constraints and the conflict between action continuity and realism caused by post-event verification of physical rationality in traditional methods by using a three-level cascaded architecture of the composite spatiotemporal arrangement analysis module and a forward dynamic verification closed loop of the physical simulation verification and iterative optimization module. Furthermore, it embeds composite spatial constraints of human geometric rigidity, collision avoidance and biomechanical limits in the generation stage, and drives iterative optimization through gradient backpropagation, thereby ensuring that the output action sequence simultaneously meets the beneficial effects of the triple standards of narrative logic temporality, physical motion realism and human physiological safety. Attached Figure Description

[0049] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Please see Figure 1 A text-driven method for automatically generating AIGC character action sequences includes the following steps:

[0052] Step 1: System Generation: Establish an AIGC character action sequence automatic generation system. The system includes a source text multi-dimensional feature acquisition module, an action semantic standardization translation module, a composite spatiotemporal arrangement analysis module, an action decoding generation module, an action sequence post-processing and verification module, and an action sequence generation output module. Through multi-module cascaded architecture design, end-to-end data flow integration, and modular interface standardization, the system achieves fully automated generation of the entire process from text semantics to character action sequences.

[0053] Step 2, Data Acquisition: The source text multi-dimensional feature acquisition module is responsible for the multi-source aggregation of original text, motion reference data, physical parameter library and performance samples;

[0054] The source text multidimensional feature acquisition module has built-in natural language text acquisition unit, motion reference data acquisition unit, physical parameter library acquisition unit and performance sample data collection unit;

[0055] The natural language text acquisition unit acquires natural language text through user input interface / API text access, specifically including action description text, action attribute modification text, spatial relationship description text, narrative scene text, emotional state text, and temporal logic text;

[0056] The motion reference data acquisition unit acquires motion reference data through motion capture equipment (OptiTrack / Vicon) and public motion databases (AMASS / Human3.6M), specifically including skeletal joint coordinate sequences, kinematic parameter sequences, dynamic parameter sequences, motion category annotation data, fine motion reference data, and multi-style motion samples.

[0057] The physical parameter library acquisition unit obtains the physical parameter library through simulation sampling by the physics engine (PhysX / MuJoCo) and calibration with human biomechanics literature. Specifically, it includes human inertial parameters, joint kinematic constraint parameters, biomechanical limit parameters, contact and collision parameters, muscle mechanical parameters and environmental physical parameters.

[0058] The performance sample data collection unit collects performance sample data through film and animation datasets (such as the Mixamo performance library) and professional actor motion capture projects. Specifically, it includes emotion-annotated action sequences, character state-action mapping pairs, narrative scene-action association samples, performance detail tension samples, stylized action samples, multi-character interactive performance samples, and improvisational performance samples.

[0059] The advantages are: by using parallel acquisition of multi-source heterogeneous data and a unified feature representation framework, it is conducive to bridging the semantic gap between four types of heterogeneous data: natural language, motion capture, physical simulation and performing arts. It enables the fusion of multi-dimensional information such as text semantics, kinematic laws, physical constraints and performance style, thereby providing complete data support and rich style references for subsequent semantic translation and motion generation.

[0060] Step 3, Semantic Translation: The action semantic standardization translation module receives multi-source aggregated data and performs standardized translation;

[0061] The action semantic standardization translation module translates feature extraction data through a six-dimensional semantic decomposition and a two-branch decoupling mapping method, specifically including the following process:

[0062] S1.1 Based on a pre-set fine-grained action lexicon and annotation rules, the original text input to the source text multi-dimensional feature acquisition module is decomposed into six dimensions of fine-grained features: action subject, core posture, motion trajectory, speed level, intensity level, and motion range. Then, the fuzzy natural language description is structurally completed and standardized through semantic completion and contextual reasoning technology based on a large language model (the implementation process of semantic completion and contextual reasoning technology based on a large language model is as follows: using a pre-trained narrative large model to analyze the implicit scene background, character identity, and behavioral motivation of the text, combining semantic matching and synonym expansion of the fine-grained action lexicon, contextual reasoning is used to complete the action dimensions not explicitly described in the text, and the completed multi-dimensional features are converted into standardized structured parameter vectors according to the annotation rules, realizing the automatic adaptation from fuzzy natural language to precise quantitative action feature parameters).

[0063] S1.2 Then, a fine-grained semantic encoder is used to extract specific features from the decomposed multi-dimensional semantic features, and relying on a self-built text-action parameter mapping knowledge base, the text description is automatically mapped into accurately quantified basic action parameters.

[0064] S1.3 Finally, a semantic dual-branch decoupled architecture is adopted to separate action features into basic action ontology features and action attribute regulation features. The basic branch identifies core action types such as waving, walking, and raising hands, while the attribute branch independently parses semantic modifiers such as force, gentleness, speed, and slowness. Finally, the quantitative regulation parameters of force, speed, amplitude, and duration are output, completing the standardized translation of natural language text into structured and decoupled action feature parameters.

[0065] The advantages are: through the six-dimensional semantic decomposition and dual-branch decoupling mapping mechanism, it is beneficial to accurately convert fuzzy natural language descriptions into structured and quantified action feature parameters, realize the independent parsing and flexible combination of action ontology type and attribute modification semantics, thereby improving the translation accuracy and controllability of text semantics to action parameters.

[0066] Step 4, Composite Spatiotemporal Choreography: The composite spatiotemporal choreography analysis module is responsible for optimizing the composite choreography under spatiotemporal constraints of the standardized translated motion feature parameters, and outputting the optimal motion sequence parameters that conform to physical laws and narrative logic;

[0067] The composite spatiotemporal orchestration analysis module incorporates a multi-branch parallel attention analysis unit, a temporal priority analysis unit, and a spatial feature constraint loss analysis unit.

[0068] The multi-branch parallel attention analysis unit calculates global multimodal context feature weights based on different modal features of text semantics, action pose, and temporal position, combined with an adaptive weighted fusion algorithm. The calculation formula is as follows:

[0069]

[0070] In the formula, m represents the number of parallel attention branches. These represent the query, key, and value matrices, derived from input feature encodings of different modalities. The matrix representing the query, key, and value of the i-th branch is derived from the input feature encoding of different modalities. This represents the attention calculation function for the i-th branch. Represents the adaptive fusion weights of the i-th branch (generated through a dynamic routing mechanism with learnable parameters or input dependencies);

[0071] Based on the calculated global multimodal context feature weights The attention outputs of each branch are adaptively weighted and fused to generate a joint multimodal feature tensor that integrates global context semantics, pose structure and temporal position information. This tensor is then passed to the temporal priority analysis unit as the input basis for inter-frame dependency modeling.

[0072] The advantage is that it calculates global multimodal context feature weights. This is beneficial for dynamically capturing deep correlations and complementary information between different modal features, thereby overcoming the one-sidedness of single modal feature expression, generating joint multimodal feature representations that integrate global contextual semantics, pose structure and temporal position information, and ultimately providing a highly robust feature input foundation for subsequent temporal orchestration.

[0073] The temporal priority analysis unit calculates the long and short-term temporal dependencies between action frames and the inter-frame priority ranking based on the long and short-term temporal dependencies between different frames in the action sequence, combined with a priority constraint ranking algorithm. The calculation formula is as follows:

[0074]

[0075] In the formula, Indicates the first The final temporal priority score of the frame action, the higher the value, the higher the execution priority of the frame in subsequent sequence arrangement and sampling; , These represent the short-term independent contribution coefficient and the long-term dependent contribution coefficient, respectively, which are used to balance the weight of the influence of the current frame’s own features and the association with historical frames on the priority. This represents the Sigmoid activation function, which maps short-term features to the (0,1) interval; , These represent the learnable weight matrix and bias term of the short-term feature linear mapping, respectively, used to extract the instantaneous temporal saliency of the current frame; Represents the motion feature vector of frame t (the fused feature from the output of the normalized translation module); This indicates the length of the context window (the number of historical frames covered) for long-range dependency modeling. Represents the time-series decay factor ( ∈(0,1), controls the exponentially decreasing rate at which the priority of the current frame is affected by the farther away frame; This represents the cosine similarity function, used to quantify the features of the current frame. Features of historical frames The degree of similarity in their postures and movement trends;

[0076] According to the Final timing priority score for frame actions The action sequence is sorted in descending order and keyframes are extracted to determine the temporal anchor points of high-priority action frames. The frame sampling density and interpolation strategy are dynamically adjusted based on the score distribution.

[0077] When the Final timing priority score for frame actions When ∈(0.7, 1.0] (high priority interval): mark the frame as a core key frame, preserve its motion pose data completely, do not downsample, and use it as a time anchor point for bidirectional interpolation reference to ensure accurate reproduction and coherent transition of the frame in the sequence arrangement;

[0078] When the Final timing priority score for frame actions When ∈ [0.3, 0.7] (medium priority interval): perform adaptive sampling and linear interpolation on the frame, based on... The specific values ​​are retained proportionally for a portion of the frame ( The closer to 0.7, the higher the retention rate; the closer to 0.3, the lower the retention rate. Employing keyframes before and after the frame for pose transition interpolation to maintain motion smoothness, thereby reducing data redundancy while ensuring sequence fluency.

[0079] When the Final timing priority score for frame actions When ∈[0, 0.3) (low priority interval): the frame is determined to be a temporally redundant frame and is subjected to sparsification processing (only retaining the action trend features). The transition posture is generated by using the preceding and following high / medium priority frames to perform Bézier curve or spline interpolation fitting. The sequence length is significantly compressed without affecting the semantic expression of the action, thereby improving the generation efficiency and the sense of rhythm of the movement, thus providing a priority scheduling basis for the rearrangement and rhythm arrangement of subsequent action sequences.

[0080] The advantage is: by calculating the first... Final timing priority score for frame actions This helps to distinguish between core key frames, transition frames, and redundant frames in an action sequence, thereby enabling dynamic adjustment of frame sampling density and intelligent selection of interpolation strategies. It compresses the amount of sequence data while ensuring the complete expression of action semantics and motion coherence, and can also improve generation efficiency and enhance the rhythm and narrative tension of the action.

[0081] The spatial feature constraint loss analysis unit calculates the spatial constraint loss value of the human skeleton's geometric structure, physical collision, and biomechanical limits based on the topological rigidity constraint of the human skeleton, the spatial penetration potential energy field, and the physiological limit function of joint angular momentum. This ensures the rationality and realism of the movement space. The calculation formula is as follows:

[0082]

[0083] The calculation formulas for each sub-item are as follows:

[0084]

[0085]

[0086]

[0087] In the formula, This represents the total spatial constraint loss value, used for backpropagation optimization of the generated action sequence; , , These represent the balance weighting coefficients for geometric stiffness loss, collision penetration loss, and angular momentum limit loss, respectively. The geometric stiffness loss term represents the constraint that keeps the bone length constant, preventing joint dislocation or tensile deformation; where... For the skeleton topological edge set, This represents the total number of edges of the skeleton. , Let be the three-dimensional spatial coordinates of the a-th and b-th joints, respectively. This indicates the standard physiological length of the bone. The term represents the spatial penetration loss, which can constrain non-physical overlap between different parts of the human body; among which... Where ϵ is the total number of joints, and ϵ is the set minimum safe distance threshold. , Penalties are only imposed on penetrations that are less than the safe distance; The term representing the physiological limit of joint angular momentum loss can constrain the range of joint motion to within the limits of human biomechanics; among which... The total number of movable joints, This represents the current rotation angle (or angular momentum value) of the k-th joint. , These represent the maximum and minimum physiologically permissible angles of the joint as specified in biomechanical literature;

[0088] Based on the total loss value of spatial constraints The generated joint coordinates and rotation parameters are updated by gradient optimization through backpropagation algorithm, and iterated round by round until the loss converges to below the preset threshold. Finally, the output is a spatially reasonable action posture sequence that conforms to the geometric rigidity of the human body, has no physical penetration, and whose joint movement is within the physiological safety range.

[0089] The advantages are: by calculating the combined constraint loss of the geometric rigidity of the human skeleton, the spatial penetration potential energy and the physiological limit of joint angular momentum, it is beneficial to embed physical laws and biomechanical constraints in the early stage of motion generation, thereby preventing the generation of unrealistic phenomena such as joint dislocation, limb penetration and movement beyond the physiological range, and ultimately ensuring the rationality and authenticity of the output motion sequence in the three dimensions of human geometric structure, physical collision and physiological safety.

[0090] Step 5, Motion Decoding and Generation: The motion decoding and generation module receives the optimal motion sequence parameters output by the composite spatiotemporal orchestration analysis module, and maps the feature parameters into continuous 3D skeletal joint rotation and position data through the motion decoder (such as a diffusion model or generative adversarial network) to generate a preliminary motion sequence.

[0091] Action decoding and generation module workflow:

[0092] S2.1 Parameter Input and Preprocessing: Receives the optimal action sequence parameters (including weights of fused global multimodal context features) output by the composite spatiotemporal orchestration analysis module. Joint feature tensor and temporal priority score of each frame (Including the joint coordinates and rotation parameters after spatial constraint optimization), the input parameters are dimensionally aligned, normalized, and batch-packaged to form a standard feature vector sequence adapted to the decoder input format;

[0093] S2.2, Skeletal Mapping: The preprocessed standard feature vector sequence is input into the action decoder (based on the diffusion model or Transformer generator architecture). Through progressive denoising or autoregressive iterative deduction, the high-level action semantic features are mapped frame by frame to the rotation parameters (Euler angles or quaternions) and root node displacement parameters of the low-level 3D human skeleton joints, generating continuous frame skeleton pose data.

[0094] S2.3, Timing Optimization: The preliminary skeleton pose sequence output by the decoder is smoothed by a local timing window. The moving average filter or Kalman filter is used to eliminate high-frequency jitter between frames. The timing anchor point position output by the composite spatiotemporal arrangement analysis module is combined to verify the trajectory continuity, ensuring that the pose of high-priority key frames is accurately preserved and the low-priority transition frames are smoothly connected, forming a preliminary complete motion sequence.

[0095] S2.4, Preliminary motion sequence output: The preliminary motion sequence after temporal smoothing and trajectory optimization is formatted and packaged according to a unified data structure (including timestamps of each frame, rotation values ​​of each joint, displacement trajectory of the root node, and semantic tags of the motion), and transmitted together with intermediate process feature data to the motion sequence post-processing and verification module.

[0096] The advantages are: by using a diffusion model or generative adversarial network action decoder architecture and progressive feature mapping mechanism, it is beneficial to efficiently convert high-level semantic feature parameters into continuous and smooth 3D skeletal joint motion data. Combined with temporal anchor point verification and sliding window smoothing, it overcomes the problems of inter-frame jitter and motion incoherence caused by traditional direct mapping methods, and finally generates a preliminary action sequence that conforms to the laws of human kinematics.

[0097] Step 6, Post-processing and Verification: The motion sequence post-processing and verification module performs motion smoothing filtering, foot landing constraint correction and stylistic detail enhancement on the preliminary motion sequence, and completes text-action semantic consistency verification, physical rationality verification and biomechanical safety verification.

[0098] Action sequence post-processing and verification module workflow:

[0099] S3.1 Motion smoothing and artifact removal: Global motion smoothing enhancement is performed on the initial action sequence, Gaussian filtering is used to eliminate abrupt changes in bone joint angles and limb shaking, penetration correction and slip suppression are performed on the foot landing frames, and the joint angular velocity and angular acceleration are limited to eliminate non-physical motion artifacts.

[0100] S3.2 Tension Injection: Based on the stylized action samples and performance detail tension samples provided by the performance sample data collection unit in the source text multidimensional feature acquisition module, the motion style transfer of the action sequence to be output is performed. By adjusting the joint movement amplitude, speed curve sharpness and action frame duration, the performance tension and character personality expression are enhanced.

[0101] S3.3 Multi-dimensional verification: Perform triple verification on the enhanced action sequence: (a) Semantic consistency verification: compare the semantic matching degree between the generated action and the original text description to ensure the accurate correspondence between the action category, attribute modification and narrative scene; (b) Physical rationality verification: verify whether the rigid constraints of the skeleton topology, the self-collision penetration index and the direction of the ground contact reaction force meet the simulation standards of the physics engine; (c) Biomechanical safety verification: confirm that the rotation angle, angular momentum and muscle mechanical parameters of each joint are within the range of human physiological safety threshold.

[0102] S3.4 Report Generation: Mark the action sequences that pass all verification items as the final qualified data, summarize and generate a verification report (including quantitative scores for each dimension, correction records for non-compliant items, and final confidence assessment), and transmit the qualified action sequences along with the verification report to the action sequence generation output module.

[0103] The advantages are: motion smoothing filtering and artifact elimination of abrupt changes in bone angle and foot slippage; performance detail tension injection to achieve motion style transfer and personality expression enhancement based on style samples; multi-dimensional verification and triple validation to ensure that semantic matching, physical rationality and biomechanical safety all meet the standards; and the generation of verification reports to achieve a fully traceable quality record. This is conducive to comprehensively improving the naturalness, performance appeal and reliability of generated actions, and finally forming a closed-loop quality control system of generation-optimization-verification to ensure that the output results have sufficient accuracy and credibility in practical applications.

[0104] Step 7, Final Output: The motion sequence generation output module will encapsulate and export the verified final motion sequence in a standard motion data format (such as BVH, FBX or custom skeletal animation format), completing the entire process of automatic generation of text-driven AIGC character motion sequences.

[0105] Action sequence generation and output module workflow:

[0106] S4.1 Output Format Selection and Parameter Adaptation: Based on user presets or downstream application requirements, determine the target output format of the final motion sequence (such as BVH, FBX, BVA or custom skeletal animation JSON / binary format), automatically match the internal bone tree structure with the bone mapping table of the target format, adapt the joint naming convention, rotation representation (Euler angles / quaternions) and coordinate axis orientation convention of the target format, and automatically complete coordinate system transformation and rotation order adjustment.

[0107] S4.2 Fine-tuning of output parameters: Based on the output parameters set by the user (such as frame rate compression ratio, data precision truncation bit width, optional data channel filtering, action segment trimming range and scaling ratio), fine-tuning of action sequence data is performed to eliminate redundant data channels, reduce storage volume, and retain all key information required for complete action semantic expression.

[0108] S4.3 Data Verification and Packaging: The final motion sequence data, after format adaptation and parameter adjustment, is completely packaged according to the selected format specification, and the file header information (including bone structure description, frame rate, data channel definition and metadata tags) is written. The integrity of the packaged data is verified to ensure that the data is not damaged, the channel mapping is correct and all motion frames are completely contained.

[0109] S4.4 Action Sequence Export: Export the encapsulated standard format action sequence file to the specified storage path or push it to the downstream business system through the API interface, synchronously generate a delivery list (including file name, format type, number of frames, duration, data version and verification hash value), output the export success status indicator, and complete the closed-loop delivery of the entire process of automatic generation of AIGC role action sequences based on text-driven methods.

[0110] The advantages are: automatic mapping and coordinate system transformation from internal skeletal structure to target format is achieved through output format selection and parameter adaptation; frame rate compression, precision control and data volume reduction are achieved through fine adjustment of output parameters; data verification and packaging ensure the integrity of encapsulation and the correctness of channel mapping; and delivery list is generated and pushed to downstream business systems through action sequence export. This is conducive to opening up the data channel between AIGC action generation and downstream application scenarios such as game development, film and television production and virtual reality, so as to achieve efficient delivery and plug-and-play of the generated results, and finally form a complete closed-loop solution from text input to action sequence output.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A text-driven method for automatically generating AIGC character action sequences, characterized in that, Includes the following steps: Step 1: System Generation: Establish an AIGC character action sequence automatic generation system. The system includes a source text multi-dimensional feature acquisition module, an action semantic standardization translation module, a composite spatiotemporal arrangement analysis module, an action decoding generation module, an action sequence post-processing and verification module, and an action sequence generation output module. Step 2, Data Acquisition: The source text multi-dimensional feature acquisition module aggregates raw text, motion reference data, physical parameter library, and performance samples from multiple sources; Step 3, Semantic Translation: The action semantic standardization translation module receives multi-source aggregated data, performs standardized translation, and outputs action feature parameters; Step 4, Composite Spatiotemporal Choreography: The composite spatiotemporal choreography analysis module performs composite choreography optimization under spatiotemporal constraints on the standardized translated motion feature parameters and outputs the optimal motion sequence parameters; Step 5, Motion Decoding and Generation: The motion decoding and generation module receives the optimal motion sequence parameters output by the composite spatiotemporal orchestration analysis module, and maps the feature parameters into continuous 3D skeletal joint rotation and position data through the motion decoder to generate a preliminary motion sequence. Step 6, Post-processing and Verification: The motion sequence post-processing and verification module performs motion smoothing filtering, foot landing constraint correction and stylistic detail enhancement on the preliminary motion sequence, and completes text-action semantic consistency verification, physical rationality verification and biomechanical safety verification. Step 7, Final Output: The action sequence generation output module will encapsulate and export the verified final action sequence in a standard action data format, completing the entire process of automatic generation of AIGC character action sequences based on text-driven methods.

2. The text-driven AIGC character action sequence automatic generation method according to claim 1, characterized in that, The source text multidimensional feature acquisition module has built-in natural language text acquisition unit, action reference data acquisition unit, physical parameter library acquisition unit and performance sample data collection unit; the composite spatiotemporal choreography analysis module has built-in multi-branch parallel attention analysis unit, temporal priority analysis unit and spatial feature constraint loss analysis unit.

3. The text-driven AIGC character action sequence automatic generation method according to claim 2, characterized in that, The natural language text acquisition unit acquires natural language text through a user input interface / API text access method; The motion reference data acquisition unit acquires motion reference data through motion capture equipment and publicly available motion databases.

4. The text-driven AIGC character action sequence automatic generation method according to claim 2, characterized in that, The physical parameter library acquisition unit acquires the physical parameter library through physical engine simulation sampling and human biomechanics literature calibration; the performance sample data collection unit collects performance sample data through film and animation datasets and professional actor motion capture special dynamics.

5. The text-driven automatic generation method for AIGC character action sequences according to claim 1, characterized in that, The action semantic standardization translation module translates the feature extraction data using a six-dimensional semantic decomposition and dual-branch decoupling mapping method. The process includes the following steps: S1.1 Based on the preset fine-grained action lexicon and annotation rules, the input original text of the source text multi-dimensional feature acquisition module is decomposed into six fine-grained features: action subject, core posture, motion trajectory, speed level, force level and motion range. The fuzzy natural language description is then structurally completed and standardized through semantic completion and contextual reasoning technology based on a large language model. S1.2 Then, a fine-grained semantic encoder is used to extract specific features from the decomposed multi-dimensional semantic features, and relying on a self-built text-action parameter mapping knowledge base, the text description is automatically mapped into accurately quantified basic action parameters. S1.3 Finally, a semantic dual-branch decoupling architecture is adopted to separate action features into basic action ontology features and action attribute regulation features: (1) Basic branch recognition of core movement types such as waving, raising hands, and walking: (2) The attribute branches independently parse the semantics of the modifiers "forceful", "gentle", "fast" and "slow"; S1.4 Finally, output the quantitative control parameters of force, speed, amplitude, and duration, and complete the standardized translation of natural language text into structured and decoupled motion feature parameters.

6. The text-driven AIGC character action sequence automatic generation method according to claim 2, characterized in that, The multi-branch parallel attention analysis unit calculates global multimodal context feature weights based on different modal features of text semantics, action posture, and temporal position, combined with an adaptive weighted fusion algorithm. The calculation formula is as follows: In the formula, m represents the number of parallel attention branches. These represent the query matrix, This represents the query for the i-th branch. This represents the attention calculation function for the i-th branch. This represents the adaptive fusion weight of the i-th branch; The multi-branch parallel attention analysis unit calculates the global multimodal context feature weights. The attention outputs of each branch are adaptively weighted and fused to generate a joint multimodal feature tensor that integrates global context semantics, pose structure and temporal position information.

7. The text-driven AIGC character action sequence automatic generation method according to claim 1, characterized in that, The temporal priority analysis unit calculates the long and short-term temporal dependencies between action frames and the inter-frame priority ranking based on the long and short-term temporal dependencies between different frames in the action sequence, combined with a priority constraint ranking algorithm. The calculation formula is as follows: In the formula, Indicates the first The final timing priority score of frame actions. , These represent the short-term independent contribution coefficient and the long-term dependent contribution coefficient, respectively. This represents the Sigmoid activation function. , Let represent the learnable weight matrix and bias term of the short-term feature linear mapping, respectively. This represents the action feature vector of frame t. This indicates the length of the context window for long-range dependency modeling. Represents the time-series decay factor. This represents the cosine similarity function, used to evaluate the features of the current frame. Features of historical frames The degree of similarity in their movements, postures, and trends.

8. The text-driven AIGC character action sequence automatic generation method according to claim 7, characterized in that, The timing priority analysis unit is based on the first Final timing priority score for frame actions Value distribution dynamically adjusts frame sampling density and interpolation strategy: When the Final timing priority score for frame actions When ∈(0.7, 1.0]: mark the frame as a core keyframe, preserve its motion pose data completely, and do not perform downsampling; When the Final timing priority score for frame actions When ∈ [0.3, 0.7]: Adaptive sampling and linear interpolation are performed on the frame, based on... The specific values ​​are retained proportionally in a portion of the frames; When the Final timing priority score for frame actions When ∈[0, 0.3): the frame is determined to be a temporally redundant frame, and sparsification is performed. The transition posture is generated by Bézier curve / spline interpolation fitting using the preceding and following high / medium priority frames.

9. The text-driven automatic generation method for AIGC character action sequences according to claim 1, characterized in that, The spatial feature constraint loss analysis unit calculates the spatial constraint loss value of the human skeleton's geometric structure, physical collision, and biomechanical limits based on the topological rigidity constraint of the human skeleton, the spatial penetration potential energy field, and the physiological limit function of joint angular momentum. The calculation formula is as follows: The calculation formulas for each sub-item are as follows: In the formula, This represents the total loss value due to spatial constraints. , , These represent the balance weighting coefficients for geometric stiffness loss, collision penetration loss, and angular momentum limit loss, respectively. The geometric stiffness loss term is used to constrain the bone length to be constant, where... For the skeleton topological edge set, This represents the total number of edges of the skeleton. , Let be the three-dimensional spatial coordinates of the a-th and b-th joints, respectively. This indicates the standard physiological length of the bone. Represents the spatial penetration loss term, where Where ϵ is the total number of joints, and ϵ is the set minimum safe distance threshold. , Penalties are only imposed on penetrations that are less than the safe distance. The term representing the physiological limit of joint angular momentum loss is given by the expression, where The total number of movable joints, This represents the current rotation angle of the k-th joint. , These represent the maximum and minimum physiologically permissible angles of the joint as specified in biomechanical literature.

10. The text-driven AIGC character action sequence automatic generation method according to claim 9, characterized in that, The spatial feature constraint loss analysis unit is based on the total spatial constraint loss value. The generated joint coordinates and rotation parameters are updated by gradient optimization through backpropagation algorithm, and iterated round by round until the loss converges to below the preset threshold, and finally outputs a spatially reasonable sequence of motion postures.

Images