Digital product development-oriented data asset-driven digital human performance adjustment method

Abstract

The application belongs to the technical field of digital human animation, and discloses a data asset driven digital human performance adjustment method for digital product development. Through standardized disassembly and regular processing of original digital human performance data, four types of independent primitive units, i.e. face deformation primitives, skeleton action primitives, mouth shape synchronization primitives and interactive micro-reaction primitives, are formed. The face deformation primitives include Blendshape basic targets and composite expression key frame sequences, and the skeleton action primitives include whole body skeleton rotation key frame segments and local gesture degree of freedom parameters. An animation semantic topology label system covering animation parameter types, kinematics attributes, semantic attributes, style attributes and compatible attributes is established, a unified animation asset coding specification is matched, and a unique asset identifier is allocated to each primitive. Layered storage management is realized according to three levels of granularity, i.e. primitive layer, sequence layer and segment layer, and a parameter manifold space coordinate and interpolation compatibility degree are calibrated for each primitive.

Description

Technical Field

[0001] This invention belongs to the field of digital human animation technology, specifically a data asset-driven digital human performance adjustment method for digital product development. Background Technology

[0002] Computer animation technology is the core technology supporting the generation and adjustment of digital human performance. Currently, the industry mainly uses end-to-end generation solutions based on deep learning for adjusting digital human performance. These solutions train generation models on large-scale data and directly generate digital human expressions and movements driven by inputs such as text and speech. This approach is highly automated and efficient, but it suffers from weak controllability and poor style consistency. After the model is trained, it is difficult to quickly adapt to the customized tone requirements of different digital products. At the same time, the capabilities of the trained model are fixed, and existing digital human performance data cannot be accumulated as independently callable, combinable, and reusable data assets.

[0003] Although general-purpose 3D asset storage and distribution solutions such as Unity AssetBundle and Unreal Marketplace have emerged in the industry, enabling the engineering packaging, cross-project calling, and commercial distribution of animation resources, these solutions are all geared towards general 3D asset design and have not established a specialized governance system for the characteristics of digital human animation assets. They have obvious application boundaries and technical limitations: Unity AssetBundle only provides file-level resource packaging, dynamic loading, and hot updating capabilities, and has not established a semantic tagging system and parameter attribute labeling mechanism for digital human animation. It does not support the compatibility assessment and interpolation fusion of primitive-level animations, and cannot support the flexible combination and reuse of fine-grained animation assets. Unreal Marketplace is positioned as a platform for the trading and distribution of finished animation resources. The asset granularity within the platform is generally coarse, and digital human animations produced by different manufacturers follow their own production standards. The bone naming rules, Blendshape index definitions, and parameter output formats are inconsistent, and cross-project applications still require a lot of manual adaptation and adjustment, making it difficult to achieve efficient asset accumulation and reuse.

[0004] The existing technology still has the following core technical problems: The performance data of digital humans lacks a systematic asset governance capability. The data is scattered across different projects, and the data format and skeleton binding specifications of each project are inconsistent. The standardization level is low and the special tag system is missing, making it difficult to achieve direct reuse across scenarios and projects. Performance data needs to be collected and produced again for each new project, resulting in a duplication and waste of data resources and R&D costs. The performance tuning process is deeply tied to the model. Customization needs must be achieved through model fine-tuning and retraining. The tuning process relies on the participation of professional algorithm personnel. The development cycle is long and the technical threshold is high. It is difficult to respond to the business side's needs for rapid deployment and rapid optimization, and it cannot support the rapid iteration and agile development of digital products. Without a closed-loop optimization mechanism, user feedback and performance data during the operation of digital products cannot feed back into the iteration of digital human performance capabilities. User interaction data and animation generation links are disconnected, performance data cannot be directed to optimize specific performance units, and data assets cannot achieve continuous value appreciation.

[0005] To address the shortcomings of the existing technologies, this invention provides a method for managing and adjusting the performance of digital human animation assets that can be quickly reused in different business scenarios. Based on data asset governance, it achieves low-cost, high-efficiency customized adjustment of digital human performance through a complete chain of standardized primitive decomposition, scene constraint graph matching, baseline model assetization interface adaptation, and closed-loop iterative optimization, adapting to the business development needs of different digital products. Summary of the Invention

[0006] The purpose of this invention is to provide a data asset-driven digital human performance adjustment method for digital product development, in order to solve one or more of the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a data asset-driven digital human performance adjustment method for digital product development, comprising an asset pool construction stage, a graph construction stage, a model construction stage, a requirement analysis stage, a parameter generation stage, an output stage, and an optimization stage; Furthermore, during the asset pool construction phase, raw digital human performance data at the animation primitive level is collected. After being decomposed and standardized, the collected raw digital human performance data is divided into four types of primitive units: facial deformation primitives, skeletal motion primitives, lip-sync primitives, and interactive micro-response primitives. The facial deformation primitives include Blendshape basic targets and composite expression keyframe sequences. The skeletal motion primitives include full-body skeletal rotation keyframe segments and local gesture degree-of-freedom parameters. The lip-sync primitives include facial weight temporal sequences corresponding to phonemes. The interactive micro-response primitives include single-action keyframe units. An animation semantic topology tagging system is established to annotate all primitive units. The annotation dimensions include animation parameter type, kinematic attributes, semantic attributes, style attributes, and compatibility attributes. A unified animation asset coding standard is established simultaneously, and a unique asset identifier is assigned to each primitive. The primitive layer, sequence layer, and fragment layer are stored in a three-level granularity. The coding specification adopts a segmented and combined structure. The coding fields sequentially include the primitive category identifier, scene affiliation identifier, style attribute identifier, generation time sequence number and hierarchical code. Each segment of coding fields corresponds one-to-one with the previously labeled tag system. Encoding enables rapid identification of primitive types, applicable scenarios, and storage granularity, achieving not only unique identification of all assets but also directly supporting rapid grouping, retrieval, and batch management of asset pools by category, scenario, and style.

[0008] For each primitive, the coordinates of the parameter manifold space and the interpolation compatibility are calibrated, the fusion boundaries of different primitives in the unified parameter space are defined, and an animation primitive asset pool is constructed.

[0009] Furthermore, in the graph construction phase, typical application scenarios of digital products are sorted out, and core feature vectors of scenarios are extracted, including four dimensions: interaction type, user group attributes, product business tone, and core business objectives. Based on animation psychology and sports biomechanics theory, a quantitative mapping relationship between scenario features and animation parameter space is established, and the constraint boundaries of digital human animation under different scenarios are clarified. The constraint boundaries include the value range of facial deformation weight, the threshold of rotational degrees of freedom of skeletal joints, the slope range of keyframe interpolation curve, the amplitude scaling factor of lip-sync, and the time delay window of motion feedback. Based on the mapping relationship, a scene feature and animation parameter constraint association graph is constructed. The graph nodes are divided into scene feature nodes, animation parameter nodes, and constraint rule nodes. The edges between nodes correspond to the constraint strength coefficients. Based on the aforementioned association graph, parameter adaptation scores and interpolation weight baselines for each animation primitive in the asset pool are calculated for multiple scenarios. The available parameter boundaries of the primitive in the corresponding scenario are simultaneously calibrated to form the primitive call baseline configuration for different business scenarios.

[0010] The formula for calculating the scene adaptability score of primitive elements is as follows: This represents the scene adaptation score of the j-th animation primitive in the s-th scene, with a value range of [0, 1]. The higher the score, the better the adaptation. This represents the total number of scene parameter constraint dimensions, which is fixed at 5 dimensions, corresponding to facial deformation weight, bone rotation degrees of freedom, keyframe interpolation slope, lip-sync amplitude, and motion feedback latency, respectively. This represents the weight coefficient of the k-th constraint dimension, presupposed based on animation psychology and sports biomechanics theories, satisfying... =1; This represents the range of intrinsic parameter values ​​for the j-th primitive in the k-th parameter dimension; This represents the constraint boundary interval of the s-th scene in the k-th parameter dimension; This function represents the interval length calculation function. This represents the length of the intersection between the primitive parameter range and the scene constraint range. This represents the total length of the scene constraint interval.

[0011] Furthermore, in the model building stage, a multimodal input-driven animation parameter generation baseline model is constructed. The model sets three input branches: speech features, text semantic features, and interactive trigger signals. After cross-modal attention fusion and decoding, it outputs three types of animation parameter sequences that can directly drive the rendering engine. The animation parameter sequences are the face blendshape weight sequence, the skeleton rotation keyframe sequence, and the lip-sync parameter sequence. An animation primitive adaptation interface is designed in the feature decoding layer of the model, which corresponds one-to-one with the three parameter branches of face, action and mouth shape. It supports the dynamic injection of feature distribution parameters of the corresponding primitive. The interface has a built-in distribution alignment module. After the primitive features are injected, the distribution alignment module aligns the original parameter distribution generated by the model to the parameter space of the injected primitive, so that the style and motion features of the output animation are consistent with the primitive assets. After completing the general pre-training of the baseline model, a standardized asset call and parameter output interface is encapsulated. The model only learns the temporal coherence and multimodal alignment logic of the animation parameters, and all performance styles and semantic features are defined by the injected animation primitive assets.

[0012] Furthermore, in the requirement analysis phase, a perception module is deployed on the digital product client. The perception module is used to collect user interaction data including voice semantic text, touch operation behavior, and page dwell time, as well as product business status data including current page node, business process stage, and trigger event type. The collected real-time data is processed by feature extraction and normalization to generate a demand vector for the current scene. The demand vector is then input into the scene feature and animation parameter constraint association graph for matching and reasoning. The set of hard constraints generated by the animation parameters is obtained through parsing, including the deformation weight anchor point value corresponding to the target emotion, the upper limit of bone rotation corresponding to the movement amplitude, the keyframe density corresponding to the feedback delay, and the lip-sync scaling coefficient for voice synchronization. The constraint priority weights of the three types of parameter branches are output synchronously.

[0013] Furthermore, in the parameter generation stage, based on the target parameter constraints obtained from the interaction requirements analysis, and using semantic matching degree, parameter compatibility degree, and scene adaptability degree as comprehensive retrieval indicators, an approximate nearest neighbor retrieval is performed in the animation primitive asset pool to select the top three candidate primitive units with the highest matching degree. After completing the initial ranking and screening, the system will perform a secondary verification and screening of the top three candidate primitives based on the hard constraints of the current scenario, and remove primitives whose parameter range exceeds the constraint boundary or whose interpolation compatibility does not meet the requirements; if unqualified primitives are found, primitives ranked lower will be automatically selected to make up the number of candidates.

[0014] The animation parameter manifold space weighted interpolation algorithm is adopted. First, the candidate primitive units are mapped to a unified low-dimensional animation parameter manifold space. Then, the matching weights of each primitive and the target parameter constraints are combined to perform continuous interpolation calculations in the manifold space to generate an animation parameter set that meets the requirements of the current scene. After interpolation, kinematic compliance pre-verification is performed. Based on the skeletal binding constraints of the digital human 3D model and the Blendshape deformation boundary, the legality of the parameters is verified, and parameters that exceed the limits are truncated and smoothed.

[0015] Furthermore, in the output stage, the spatiotemporal consistency of the generated animation parameter set is checked. In the time dimension, the state of the digital human animation parameters in the previous rendered frame is obtained, and the key frame parameters are smoothed by cubic Bézier curves to achieve natural transitions between frames. At the same time, the deformation weights of high-frequency jitter are low-pass filtered. In the spatial dimension, based on the animation semantic alignment rules, the semantic strength consistency of three types of parameters—facial expression, body movement, and lip-sync—is checked, and the amplitude scaling correction is performed on parameter branches with deviations exceeding the threshold. During the verification process, hierarchical clipping detection is performed simultaneously, checking the collision risk of animation parameters sequentially from the bone level and mesh level, and fine-tuning the parameters of risk frames; The verified parameter sequence is output to the real-time rendering engine of the digital product through the standard animation interface, driving the digital human to complete the rendering presentation of the corresponding performance. At the same time, it provides keyframe downsampling and deformation accuracy simplification options for terminal devices with different performance levels.

[0016] Furthermore, in the optimization phase, an effect feedback collection module is deployed to collect animation technical quality indicators and business effect indicators. The technical quality indicators include inter-frame motion smoothness, animation semantic consistency, clipping anomaly rate, keyframe naturalness, and rendering frame rate. The business effect indicators include interaction conversion rate and satisfaction score. A comprehensive quality assessment model for animation performance is constructed, with technical quality indicators as the core weights. The effect of this primitive call and parameter generation is quantitatively scored, and the adaptation score and interpolation weight of the corresponding animation primitive under different scene constraints are dynamically adjusted according to the scoring results. Meanwhile, for animation sequences with excellent performance scores, keyframe extraction, motion normalization, and semantic topology annotation processes are used to refine them into standard animation primitive assets, which are then stored in the derivative primitive layer of the asset pool. The asset pool is periodically subjected to parametric manifold clustering for deduplication and quality grading to clean up primitive assets with low quality and low reusability.

[0017] The beneficial effects of this invention are as follows: 1. This invention standardizes and organizes raw digital human performance data to form four independent primitive units: facial deformation primitives, skeletal motion primitives, lip-sync primitives, and interactive micro-response primitives. Facial deformation primitives include Blendshape basic targets and composite expression keyframe sequences, while skeletal motion primitives include full-body skeletal rotation keyframe fragments and local gesture freedom parameters. An animation semantic topology tagging system covering animation parameter types, kinematic attributes, semantic attributes, style attributes, and compatibility attributes is established. A unified animation asset coding standard is used to assign a unique asset identifier to each primitive. Layered storage management is implemented according to a three-level granularity of primitive layer, sequence layer, and fragment layer. At the same time, parameter manifold space coordinates and interpolation compatibility are calibrated for each primitive, clarifying the fusion boundaries of different primitives in a unified parameter space. This transforms scattered performance data into standardized data assets that can be independently called and combined for reuse, avoiding repeated data collection and production work in new projects and reducing the repeated consumption of data resources and R&D costs.

[0018] 2. The multimodal input-driven animation parameter generation baseline model constructed in this invention sets up three independent input branches: speech features, text semantic features, and interactive trigger signals. After cross-modal attention fusion and decoding processing, it outputs a multi-type animation parameter sequence that can directly drive the rendering engine. The model feature decoding layer is designed with animation primitive adaptation interfaces that correspond one-to-one with the three parameter branches. The interface has a built-in distribution alignment module that can dynamically inject the feature distribution parameters of the corresponding primitives and align the original parameter distribution generated by the model to the parameter space of the injected primitives, so that the style and motion features of the output animation are highly consistent with the primitive assets. The baseline model only learns the temporal coherence and multimodal alignment logic of the animation parameters. All performance styles and semantic features are defined by the injected primitive assets. Combined with the multi-scene adaptability and available parameter boundaries calculated by the constraint graph in advance, it can directly adapt to the business tone of different digital products by calling different primitives and adjusting the constraint configuration, without the need to carry out model fine-tuning and retraining. It can quickly respond to the needs of the business side for rapid deployment and rapid optimization.

[0019] 3. This invention, through its deployed effect feedback collection module, can simultaneously collect multi-dimensional indicator data on both animation technical quality and business performance. Technical quality indicators cover dimensions such as inter-frame motion smoothness, animation semantic consistency, clipping anomaly rate, keyframe naturalness, and rendering frame rate. Business performance indicators cover dimensions such as interaction conversion rate and user satisfaction rating. Through a constructed comprehensive animation performance quality evaluation model, using technical quality indicators as core weights, it performs quantitative scoring on the effect of single primitive calls and animation parameter generation. Based on the scoring results, it adjusts the adaptation score and interpolation weight of corresponding primitives under different scene constraints, precisely optimizing the scene adaptation accuracy of primitives. Simultaneously, animation sequences with excellent performance scores are refined into standard animation primitive assets through standardized processes such as keyframe extraction, motion normalization, and semantic topology annotation, and stored in the derivative primitive layer of the asset pool. Combined with periodic parametric manifold clustering for deduplication, quality grading, and the cleanup of low-quality, low-reusability assets, it promotes continuous self-updating and quality upgrading of the asset pool, enabling business data during the operation of digital products to effectively feed back into the improvement of digital human performance capabilities. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall workflow of the method of the present invention; Figure 2 This is a flowchart of the parameter generation stage of the present invention; Figure 3 This is a diagram of the overall system architecture of the present invention. Detailed Implementation

[0021] 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.

[0022] like Figures 1 to 3 As shown, this embodiment of the invention provides a data asset-driven digital human performance adjustment method for digital product development, including an asset pool construction stage, a graph construction stage, a model construction stage, a requirement analysis stage, a parameter generation stage, an output stage, and an optimization stage. The implementation of this method is based on the following hardware and software environment configuration: On the hardware side, offline asset governance and model pre-training use computing servers equipped with NVIDIA RTX 4090 GPU (24GB GDDR6X VRAM), Intel Core i9-13900K CPU, and 64GB DDR5 memory; the real-time rendering end is divided into two types of terminals: the minimum configuration for PC is NVIDIA GTX 1660 6GB graphics card, and the minimum configuration for mobile terminal is Snapdragon 8 Gen1 / Dimensity 9000 level SoC chip.

[0023] On the software side, offline asset processing and model training are developed based on Python 3.9 and PyTorch 2.0 frameworks; the real-time rendering engine is compatible with two mainstream versions, Unreal Engine 5.1 and Unity 2022.3 LTS, and the animation parameter output follows the glTF 2.0 animation extension specification and FBX 2020 standard. Data interaction with the rendering engine is completed through the standardized Animation Stream interface, and the interface transmission frame rate supports two adaptive levels of 30fps and 60fps.

[0024] In this embodiment of the invention, during the asset pool construction phase, raw digital human performance data at the animation primitive level is collected. After being decomposed and standardized, the collected raw digital human performance data is divided into four types of primitive units: facial deformation primitives, skeletal motion primitives, lip-sync primitives, and interactive micro-response primitives. The facial deformation primitives include Blendshape basic targets and composite expression keyframe sequences. The skeletal motion primitives include full-body skeletal rotation keyframe segments and local gesture freedom parameters. The lip-sync primitives include facial weight temporal sequences corresponding to phonemes. The interactive micro-response primitives include single-action keyframe units. The decomposition process is divided according to data modality, human body movement parts and functional uses. First, audio-driven lip movement data, facial expression change data reflecting emotional state, whole-body skeletal movement data controlling limb posture, and discrete interactive trigger action data responding to user operations are separated from the original mixed data. After the splitting is completed, the data format is aligned using industry-standard bone naming rules, Blendshape morphological indexing rules, and global timestamp standards. Invalid frame data with stuttering, pose distortion, or no actual semantics are batch-screened and removed. Finally, the complete data stream is cut according to the independent action semantic boundaries to extract reusable primitive fragments with independent use value and the smallest granularity.

[0025] An animation semantic topology tagging system is established to annotate all primitive units. The annotation dimensions include animation parameter type, kinematic attributes, semantic attributes, style attributes, and compatibility attributes. A unified animation asset coding standard is established simultaneously, and a unique asset identifier is assigned to each primitive. The primitive layer, sequence layer, and fragment layer are stored in a three-level granularity. Each annotation dimension is managed using a hierarchical tree-like tag structure. The animation parameter type dimension has a first-level tag based on the major category to which the primitive belongs, quickly distinguishing four main categories: face, skeleton, lip shape, and micro-reactions. The kinematic attribute dimension details physical motion characteristic tags such as motion amplitude, speed, and joint range of freedom. The semantic attribute dimension is divided into multiple layers of semantic tags, including emotion categories, action expression intentions, and human-computer interaction social meanings. The style attribute dimension distinguishes the overall tone of the digital human's performance, the age range of users it suits, and the style of applicable product scenarios. The compatibility attribute dimension clearly defines the types of primitives that can be linked and the allowed range of parameter changes when they are combined. All tags are entered into the system using fixed enumeration values ​​and are permanently bound to the unique identifier of each primitive.

[0026] The three-tiered storage hierarchy is divided in descending order of data granularity. The primitive layer stores animation units with complete and independent semantics, corresponding to a complete set of emotional expressions, coherent actions, or standard interaction flows, and is the main unit for external calls. The sequence layer is responsible for splitting the complete primitives in chronological order and storing the continuous keyframe subsequences after splitting, each corresponding to a specific action stage or a specific expression change process. The fragment layer is the smallest storage unit in the entire asset pool, mainly storing single-frame images or temporal parameters within a very short time, corresponding to a specific fixed facial deformation state or a single skeletal pose. The three-tiered storage units establish hierarchical relationships through unified asset coding.

[0027] For each primitive, the coordinates of the parameter manifold space and the interpolation compatibility are calibrated, the fusion boundaries of different primitives in the unified parameter space are defined, and an animation primitive asset pool is constructed.

[0028] The interpolation compatibility is calibrated by comprehensively referring to two core indicators: the range of overlap of parameter dimensions between different primitives, the overlap of kinematic constraints, and the degree of semantic expression matching. Finally, a standardized quantitative level is formed. The higher the compatibility level, the lower the probability of abnormal problems such as skeletal distortion and facial expression disorder after multiple primitives are fused. The fusion boundary is defined by combining the maximum and minimum ranges of the primitive's own parameters, as well as the neighborhood distance of the primitive in the overall parameter manifold space. The system will automatically exclude primitive combinations that exceed the fusion boundary from the interpolation candidate range.

[0029] In this embodiment of the invention, during the graph construction stage, typical application scenarios of digital products are sorted out, and core feature vectors of the scenarios are extracted, including four dimensions: interaction type, user group attributes, product business tone, and core business objectives. Based on animation psychology and sports biomechanics theory, a quantitative mapping relationship between scenario features and animation parameter space is established, and the constraint boundaries of digital human animation under different scenarios are clarified. The constraint boundaries include the value range of facial deformation weight, the threshold of rotational degrees of freedom of skeletal joints, the slope range of keyframe interpolation curves, the amplitude scaling factor of lip-sync, and the time delay window of motion feedback. Based on the mapping relationship, a scene feature and animation parameter constraint association graph is constructed. The graph nodes are divided into scene feature nodes, animation parameter nodes, and constraint rule nodes. The edges between nodes correspond to the constraint strength coefficients. The graph construction is carried out step by step according to the process of entity input, rule conversion, and relationship connection. First, all application scenario features and digital human animation operation parameters obtained from the project are entered into the system one by one as the basic entity nodes of the graph. Second, combined with animation psychology and sports biomechanics, the qualitative constraints are transformed into standardized constraint rules that can be recognized by the system, generating independent rule nodes. Finally, following the logic of scene features pointing to constraint rules, and constraint rules pointing to animation parameters, the connection between nodes is constructed; the constraint strength coefficients for the connection are set according to the rule enforcement level, with the highest strength coefficient configured for hard constraints that cannot be broken, and the gradient strength coefficient configured for soft constraints that are only used for reference and guidance.

[0030] Based on the aforementioned association graph, parameter fit scores and interpolation weight baselines are calculated for each animation primitive in the asset pool under multiple scenarios. The available parameter boundaries of the primitive in the corresponding scenario are simultaneously calibrated, forming a primitive call baseline configuration for different business scenarios. This limits the parameter search space in the real-time generation stage, realizes the pre-matching of animation assets and business scenarios, and reduces the computational overhead in the real-time adjustment stage.

[0031] When calculating the fit, the complete feature vector corresponding to a single business scenario is first input into the constraint association graph. The graph inference outputs the complete set of animation parameter constraints and restrictions for that scenario. Then, the parameter operating range and motion characteristics of a single primitive in the asset pool are compared and matched with this set of constraints item by item. The larger the overlap and the higher the fit, the higher the final scenario fit score of the primitive. The interpolation weight baseline is generated based on the fit score and converted proportionally. In the same business scenario, the primitive with better fit performance will occupy a larger weight ratio by default in the subsequent parameter interpolation process and take the lead in the final animation performance.

[0032] In this embodiment of the invention, during the model building stage, a multimodal input-driven animation parameter generation baseline model is constructed. The model is set with three input branches: speech features, text semantic features, and interactive trigger signals. After cross-modal attention fusion and decoding, it outputs three types of animation parameter sequences that can directly drive the rendering engine. The animation parameter sequences are a face blendshape weight sequence, a bone rotation keyframe sequence, and a lip-sync parameter sequence. Each of the three input branches has an independent feature extraction unit. The speech feature branch uses a multi-layer one-dimensional convolutional network with a bidirectional gated recurrent unit structure to extract the prosodic features and temporal dependency features of the audio. The text semantic feature branch uses a multi-layer lightweight Transformer coding layer to extract the semantic embedding features of the text. The interaction trigger signal branch uses a multi-layer fully connected coding layer to map discrete trigger events into continuous feature vectors of fixed dimensions.

[0033] The specific network parameter settings for each branch are as follows: The speech feature branch uses a 3-layer one-dimensional convolutional network with kernel sizes of 5, 3, and 3, and output channels of 64, 128, and 256 respectively, followed by a 1-layer bidirectional GRU module with a hidden layer dimension of 256, ultimately outputting a 256-dimensional temporal feature vector; the text semantic feature branch uses a 3-layer lightweight Transformer encoder with a single encoder having 4 self-attention heads, a hidden layer dimension of 256, and a feedforward network dimension of 512, outputting a 256-dimensional semantic embedding vector; the interaction trigger signal branch uses a 2-layer fully connected coding layer with a hidden layer dimension of 128, outputting a 256-dimensional continuous feature vector; the cross-modal attention fusion module uses 4 cross-attention heads, and after fusion, outputs a unified 256-dimensional temporal feature.

[0034] The cross-modal attention fusion module adopts a cross-attention mechanism, using speech temporal features as query vectors and text semantic features and interaction features as key-value vectors to complete the temporal alignment and deep fusion of multimodal features. The fused feature inputs share a temporal decoding layer and are then connected to three independent parameter output heads to generate animation parameter sequences for the corresponding categories.

[0035] An animation primitive adaptation interface is designed in the feature decoding layer of the model, which corresponds one-to-one with the three parameter branches of face, action and mouth shape. It supports the dynamic injection of feature distribution parameters of the corresponding primitive. The interface has a built-in distribution alignment module. After the primitive features are injected, the distribution alignment module aligns the original parameter distribution generated by the model to the parameter space of the injected primitive, so that the style and motion features of the output animation are consistent with the primitive assets. The distribution alignment module adopts the maximum mean difference alignment strategy. The pre-stored parameter mean vector, variance vector and covariance matrix of the injected primitives are used as the target distribution benchmark. The parameter distribution of the original output of the model is adjusted through a learnable linear transformation layer, so that the deviation between the statistical characteristics of the output parameters and the statistical characteristics of the target primitive parameter space is controlled within a preset range, while preserving the temporal variation trend and multimodal alignment relationship of the original output.

[0036] The optimization objective of distribution alignment is to minimize the maximum mean difference (MMD) between the generated parameter distribution and the target primitive parameter distribution. A multi-scale Gaussian kernel is used as the kernel function, with kernel bandwidths set to 0.1, 1, and 10, respectively. The linear transformation layer uses orthogonal initialization, and the learning rate for the alignment process is set to 1×10⁻⁶. -4 The iteration steps are 500 steps. After alignment, the mean deviation of the parameter distribution is controlled within 5%, and the covariance deviation is controlled within 10%, ensuring the consistency between the output style and the injected primitives.

[0037] After completing the general pre-training of the baseline model, a standardized asset call and parameter output interface is encapsulated. The model only learns the temporal coherence and multimodal alignment logic of animation parameters, without carrying any prior knowledge of style or emotion. All performance style and semantic features are defined by the injected animation primitive assets, and the model only serves as the execution carrier for generating animation parameters.

[0038] The baseline model pre-training uses a publicly available large-scale multimodal digital human animation dataset. The training is divided into two stages: the first stage is multimodal alignment pre-training, with speech-lip synchronization loss and text-expression semantic consistency loss as the core loss functions, to train the model to master the basic mapping logic between multimodal input and animation parameters; The second stage is temporal coherence pre-training, which uses inter-frame parameter smoothness loss and kinematic compliance loss as auxiliary loss functions to constrain the natural inter-frame transition of output parameters. During the pre-training process, the parameters of the feature extraction unit and the fusion module are fixed, and only the weights of the decoding layer and the parameter output head are updated. After training, all parameters of the backbone network are frozen.

[0039] During the formal operation of the model, the primitive adaptation interface adopts a dynamic hot-swappable mode, which can switch and inject different animation primitive assets in real time according to the current business scenario and constraint rules; the model backbone network remains in a fixed state throughout the process without retraining or adjusting parameters, and only changes the final animation style, emotional characteristics and action features of the digital human by changing the injected primitive assets.

[0040] In this embodiment of the invention, during the requirement analysis phase, a perception module is deployed on the digital product client. The perception module is used to collect user interaction data including voice semantic text, touch operation behavior, and page dwell time, as well as product business status data including the current page node, business process stage, and trigger event type. All user interaction data undergoes compliant anonymization processing throughout the entire collection, transmission, and storage process: voice data is only extracted locally for acoustic features and semantic text results, and the original audio files are not stored; user behavior data has its device unique identifier, user account ID, and other personally identifiable information removed, retaining only anonymized behavioral temporal features; all data storage uses AES-256 symmetric encryption and is used only for local animation parameter generation and effect quality statistics, and is not transmitted to third parties, fully complying with relevant compliance requirements for data security and personal information protection.

[0041] The collected real-time data is processed by feature extraction and normalization to generate a demand vector for the current scene. The demand vector is then input into the scene feature and animation parameter constraint association graph for matching and reasoning. The feature extraction process converts the fragmented discrete data collected by the client into feature values ​​that are consistent with the core feature vector dimension of the preset scenario. Then, through normalization processing, all values ​​are gathered into a unified range and finally combined to form a requirement vector that can fully express the current business requirements. Graph matching reasoning employs a subgraph matching approach, quickly retrieving the set of scene feature nodes with the highest overlap with the demand vector within a vast relational graph. Then, it deduces layer by layer along the connections between nodes, linking all associated constraint rule nodes and animation parameter nodes, and finally integrating and outputting the entire set of parameter constraint results.

[0042] The set of hard constraints generated by the animation parameters is obtained through parsing, including the deformation weight anchor point value corresponding to the target emotion, the upper limit of bone rotation corresponding to the movement amplitude, the keyframe density corresponding to the feedback delay, and the lip-sync scaling coefficient for voice synchronization. The constraint priority weights of the three types of parameter branches are output synchronously.

[0043] The set of hard constraints obtained from the analysis serves as an inviolable bottom line rule, and all generated animation parameters must fall within the range defined by the hard constraints. The constraint priority weights are used to distinguish the order of control of the three types of parameters: face, bones, and mouth shape. Parameter branches with higher priority have greater decision-making power in the interpolation calculation and amplitude correction process, ensuring that the core performance content is not interfered with by secondary parameters.

[0044] In this embodiment of the invention, during the parameter generation stage, based on the target parameter constraints obtained from the interaction requirements analysis, and using semantic matching degree, parameter compatibility degree, and scene adaptability degree as comprehensive retrieval indicators, an approximate nearest neighbor retrieval is performed in the animation primitive asset pool to select candidate primitive units with high matching degree ranking. The real-time collected voice, text and interaction trigger signals are input into the pre-trained multimodal animation parameters to generate a baseline model, resulting in an initial animation parameter sequence with basic temporal coherence. Using the initial parameter sequence as a retrieval benchmark, candidate primitives are retrieved from the animation primitive asset pool in combination with target parameter constraints. Finally, the final animation parameter set adapted to the current scene is generated through manifold space interpolation.

[0045] The three indicators of semantic matching degree, parameter compatibility degree and scene adaptability will be uniformly quantified into standard scores within a fixed range. Combined with the constraint priority weights output in the requirement analysis stage, the three scores will be weighted and summed to obtain the comprehensive matching score of each primitive. The retrieval process uses a vector index library pre-built in the parametric manifold space to perform an approximate nearest neighbor search. The system first performs a large-scale initial screening based on scene adaptability, filtering out primitives that obviously do not meet the requirements of the current scene and narrowing the search range. Then, it combines the comprehensive matching score to complete the sorting, and selects the high-ranking high-quality primitives as interpolation candidate units according to the score.

[0046] Formula for calculating the matching degree of animation primitive comprehensive retrieval: This represents the overall search matching score of the i-th animation primitive, with a value range of [0, 1], and is used for primitive sorting and filtering; The constraint priority weights, representing the semantic matching degree, are output from the requirements parsing phase; represents the semantic matching score between the i-th primitive and the target requirement, calculated based on the cosine similarity between the primitive semantic label and the requirement semantic vector, with a value range of [0, 1]. The constraint priority weights, representing parameter compatibility, are output from the requirements analysis phase. The i-th primitive represents the compatibility score between the i-th primitive and the target parameter constraint, calculated as the proportion of the intersection between the primitive parameter range and the set of hard constraints, with a value range of [0, 1]. The constraint priority weights, representing the scene adaptability, are output from the requirements analysis phase; Weight constraints: The weights of the three priority categories satisfy... + + =1.

[0047] The animation parameter manifold space weighted interpolation algorithm is adopted. First, the candidate primitive units are mapped to a unified low-dimensional animation parameter manifold space. Then, the matching weights of each primitive and the target parameter constraints are combined to perform continuous interpolation calculations in the manifold space to generate an animation parameter set that meets the requirements of the current scene. A unified low-dimensional parameter manifold space is pre-constructed through unsupervised dimensionality reduction. The standardized high-dimensional parameter vectors of all primitives in the asset pool are used as training samples. The t-distributed neighborhood embedding algorithm is used to map the original high-dimensional animation parameters to a fixed low-dimensional manifold space. The mapping process preserves the semantic similarity and parameter compatibility between primitives. When mapping primitives, the pre-trained mapping function is directly called to convert the parameter vectors of candidate primitives into coordinate points in the manifold space.

[0048] The dimension of the reduced manifold space is fixed at 64. The perplexity parameter of the t-SNE algorithm is set to 30, and the maximum number of iterations is 1000. After mapping, the average neighborhood distance of primitives of the same semantic category in the manifold space does not exceed 0.2, and the average distance of primitives of different semantic categories is not less than 0.8, ensuring the semantic discriminability and interpolation rationality of the manifold space.

[0049] Interpolation uses the comprehensive retrieval matching degree of each candidate primitive as the basic weight, and takes the set of hard constraints and the constraint priority weights obtained in the requirement analysis stage as the interpolation boundary conditions. The hard parameter boundary is directly used as the truncation threshold of the interpolation result, and the constraint priority weights correspond to the interpolation influence coefficients of each parameter dimension of the manifold space. The interpolation calculation adopts the spherical linear interpolation method, calculates the coordinates of the target interpolation point in the manifold space, and then restores it to the high-dimensional animation parameter set through the pre-trained inverse mapping network.

[0050] Animation parameter manifold space weighted interpolation formula: Represents the coordinates of the target interpolation point in the low-dimensional parametric manifold space; This indicates the number of candidate primitives participating in the interpolation, and the top 3 primitives with the highest overall matching degree are always selected. This represents the overall retrieval matching score of the i-th candidate primitive; The constraint priority weight of the i-th parameter dimension is obtained by mapping the branch weights of the three types of parameters (face, skeleton, and mouth shape) output in the requirements analysis phase. This represents the pre-calibrated coordinate point of the i-th candidate primitive in the unified low-dimensional parametric manifold space; This represents the final set of high-dimensional animation parameters that can directly drive the rendering engine. This represents the inverse mapping function from the pre-trained low-dimensional manifold space to the original high-dimensional animation parameter space.

[0051] After interpolation, kinematic compliance pre-verification is performed. Based on the skeleton binding constraints of the digital human 3D model and the Blendshape deformation boundary, the legality of the parameters is verified, and parameters that exceed the limits are truncated and smoothed to avoid animation anomalies such as clipping and bone distortion.

[0052] For abnormal parameters that exceed the preset constraint boundaries, truncation is performed first, directly correcting the excessive parameter values ​​to the maximum or minimum boundary values ​​allowed by the rules. After truncation, smoothing is performed simultaneously, with the key frame where the abnormality occurs as the center, and gradient parameter transition adjustments are made to the data of multiple adjacent frames before and after.

[0053] The correction process preserves the overall trend and rhythm of the original animation parameters, and only fine-tunes and optimizes the values ​​in the out-of-limit range to prevent abrupt changes in movements or expressions after parameter correction.

[0054] In this embodiment of the invention, during the output stage, the spatiotemporal consistency of the generated animation parameter set is checked. In the time dimension, the state of the digital human animation parameters of the previous rendering frame is obtained. The key frame parameters are smoothed between frames using cubic Bézier curves to achieve natural transitions between frames. At the same time, the deformation weights of high-frequency jitter are low-pass filtered to eliminate the incongruity of action jumps and facial expression changes. The smoothing and filtering operations in the time dimension do not change the semantic strength of the parameters; they only optimize the transition effect between frames. The verification work in both dimensions is performed in parallel. If an abnormal parameter is detected in either dimension, it will be marked synchronously and enter a unified correction process.

[0055] In the spatial dimension, based on the animation semantic alignment rules, the semantic strength consistency of three types of parameters—facial expression, body movement, and lip-sync—is verified. Parameter branches with deviations exceeding the threshold are corrected by amplitude scaling to avoid the problem of multimodal animation expression conflicts. The animation semantic alignment rules predefine a fixed proportional relationship between the amplitude of facial expression changes, the amplitude of body movement, and the amplitude of mouth opening and closing under different semantic intensity levels. During the verification process, the system extracts the real-time semantic intensity values ​​corresponding to the three types of parameters, namely face, body, and mouth, and judges the magnitude of the deviation by comparing them with the preset proportional relationship.

[0056] When the deviation value exceeds the system's preset threshold, the parameter branch with the most prominent current expression intention is used as the benchmark, and the other two types of parameters are scaled and adjusted proportionally.

[0057] During the verification process, hierarchical clipping detection is performed simultaneously, checking the collision risk of animation parameters sequentially from the bone level and mesh level, and fine-tuning the parameters of risk frames; The hierarchical clipping detection is performed step by step, following the order of skeleton first and then mesh. The skeleton hierarchical detection mainly calculates the shortest spatial distance between each bone structure to determine whether there is a risk of bone clipping or limb collision. If a problem is detected, the rotation angle of non-core limb bones is adjusted first to avoid collision problems without affecting the main body's movement.

[0058] Mesh-level detection targets the digital human appearance model, calculates the intersection range between model mesh vertices, determines the severity of clipping in the surface model, and completes fine-tuning by slightly shifting local deformation weights.

[0059] The verified parameter sequence is output to the real-time rendering engine of the digital product through the standard animation interface, driving the digital human to complete the rendering presentation of the corresponding performance. At the same time, it provides keyframe downsampling and deformation accuracy simplification options for terminal devices with different performance levels.

[0060] The system automatically matches the corresponding adaptation scheme according to the hardware performance and rendering capabilities of the terminal device. For high-performance terminals, it retains the complete number of keyframes and deformation accuracy; for mobile terminals and low-end terminals, it performs keyframe downsampling and deformation level simplification to reduce the amount of rendering computation without destroying the overall action, facial expression semantics and motion logic.

[0061] In this embodiment of the invention, during the optimization phase, an effect feedback collection module is deployed to collect animation technical quality indicators and business effect indicators. The technical quality indicators include inter-frame motion smoothness, animation semantic consistency, clipping anomaly rate, keyframe naturalness, and rendering frame rate. The business effect indicators include interaction conversion rate and satisfaction score. A comprehensive quality assessment model for animation performance is constructed, with technical quality indicators as the core weights. The model quantifies and scores the effect of this primitive call and parameter generation. Based on the scoring results, the model dynamically adjusts the adaptation score and interpolation weight of the corresponding animation primitive under different scene constraints, thereby optimizing the scene adaptation accuracy of the primitive. The animation performance comprehensive quality evaluation model adopts the logic of hierarchical weighting combined with fuzzy comprehensive evaluation, setting technical quality indicators as the first-level evaluation dimension and corresponding to the main weight proportion, and setting business performance indicators as the second-level evaluation dimension and corresponding to the remaining weight proportion. Sub-indicators under each dimension are weighted using the entropy weighting method. All indicators are normalized and then weighted to obtain a comprehensive quality quantification score in the range of 0 to 1. The score results are stored in conjunction with the corresponding primitive identifier and scene feature identifier.

[0062] The formula for calculating the overall quality assessment score of animation performance is as follows: This represents the overall quality score of the animation performance, with a value range of [0, 1]. The higher the score, the better the animation performance. This represents the overall weighting coefficient of the technical quality indicator. The document explicitly states that technical quality is the core weight, with a preset value range of [0.7, 0.8]. denoted as the normalized score of the p-th technical quality indicator, with a value range of [0, 1]; p = 1 to 5 correspond to inter-frame motion smoothness, animation semantic consistency, clipping anomaly rate, keyframe naturalness, and rendering frame rate, respectively. The entropy weight method is used to calculate the weight of the p-th technical quality sub-indicator, satisfying the following condition: =1; q represents the normalized score of the q-th business performance indicator, with a value range of [0, 1]; q = 1 to 2 correspond to the interaction conversion rate and user satisfaction score, respectively. The entropy weight method is used to calculate the weight of the q-th business performance sub-indicator, satisfying the following condition: =1.

[0063] The system uses incremental updates to dynamically adjust parameters and sets a unified quality pass score as the judgment standard. When the overall animation quality score is higher than the pass threshold, it means that the primitive performs well in the current scene. The system will then slightly increase the primitive's adaptability score and interpolation weight ratio in this scene according to the corresponding proportion. When the overall quality score is lower than the passing threshold, it indicates that the basic element adaptation effect is poor, and the system will reduce the corresponding score and weight proportionally. The parameter adjustment range and the degree of deviation of the score from the passing line are positively correlated, and upper and lower limits are set for the single adjustment range.

[0064] Meanwhile, for animation sequences with excellent performance scores, keyframe extraction, motion normalization, and semantic topology annotation processes are used to refine them into standard animation primitive assets, which are then stored in the derivative primitive layer of the asset pool. The asset pool is periodically subjected to parametric manifold clustering for deduplication and quality grading to clean up primitive assets with low quality and low reusability.

[0065] The asset pool will be maintained according to a fixed period. Clustering and deduplication are based on the coordinate distance of primitives in the parametric manifold space. Primitives with a coordinate distance less than a preset threshold are judged as highly similar content. The system will automatically retain the high-quality primitives with the highest historical quality scores and integrate the compatibility attributes and tag information of other similar primitives to improve the main asset content.

[0066] Quality grading is based on a comprehensive assessment of the long-term historical average quality score of primitives and the number of times they have been called and reused, classifying them into high-quality assets, regular assets, and assets that need optimization. The asset cleanup process mainly targets inefficient primitives that have no call records for multiple consecutive maintenance cycles and whose quality scores are consistently in the lowest range. Data backup will be completed before formal cleanup.

[0067] During the extraction process, high-quality animation sequences are first assessed for semantic integrity, retaining only animation clips with independent semantics and reusable functionality as candidate primitives. After keyframe extraction, motion normalization, and semantic annotation are completed, the newly generated derived primitives undergo parametric manifold space coordinate calibration, interpolation compatibility testing, and encoding allocation. Only after all processes pass verification are they formally incorporated into the corresponding derived primitive layer of the asset pool for management.

[0068] To verify the practical effectiveness of this method, a comparative test was conducted using a digital human scenario for intelligent customer service in digital financial products. The test set included 100 multimodal interaction samples with durations of 5-10 seconds. The baselines for comparison were an end-to-end generative digital human model and a general asset retrieval scheme based on Unity AssetBundle. The core test results are shown in the table below:

[0069] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0070] 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 data asset-driven digital human performance adjustment method for digital product development, characterized in that, The specific steps include the following: During the asset pool construction phase, the original performance data of digital human animation is collected, a tagging system is established to complete the primitive annotation, a unified coding standard is set and a unique identifier is assigned to each primitive, the data is stored in layers according to granularity and the parameter manifold space attributes are labeled, and the animation primitive asset pool is built. In the graph construction phase, we sort out the application scenarios of digital products and extract the core feature vectors of the scenarios, establish a quantitative mapping relationship between scenario features and animation parameter space, construct a graph of the relationship between scenario features and animation parameter constraints, calculate the primitive scenario adaptability and generate the call baseline configuration. In the model building phase, a baseline model for generating animation parameters driven by multimodal input is built. The input features are fused and decoded through cross-modal attention to output the animation parameter sequence. The primitive adaptation interface of the built-in distribution alignment module is set in the feature decoding layer, and the standardized interface is encapsulated after pre-training is completed. During the requirement analysis phase, user interaction and product business status data are collected to generate a requirement vector for the current scenario. Input constraint association graph matching and reasoning are used to obtain parameter constraints and corresponding weights. In the parameter generation stage, based on the target parameter constraints, candidate primitives are searched and screened in the animation primitive asset pool using preset comprehensive search indicators. The animation parameter set is generated using the parametric manifold space weighted interpolation algorithm, and kinematic compliance pre-verification is performed and out-of-bounds parameters are corrected. During the output stage, the animation parameter set is checked for consistency and clipping detection. The checked animation parameters are then output to the rendering engine, providing multi-terminal adaptation options. During the optimization phase, a feedback collection module is deployed, a comprehensive quality assessment model for animation performance is built and quantitative scoring is performed, primitive adaptation and interpolation weights are adjusted based on the scores, high-quality animation sequences are extracted as standard primitive assets and added to the database, and the asset pool is optimized regularly.

2. The data asset-driven digital human performance adjustment method for digital product development according to claim 1, characterized in that, During the asset pool construction phase, the collected raw digital human performance data is decomposed and standardized to form four types of primitive units: facial deformation primitives, skeletal motion primitives, lip-sync primitives, and interactive micro-response primitives. The facial deformation primitives include Blendshape basic targets and composite expression keyframe sequences; the skeletal motion primitives include full-body skeletal rotation keyframe segments and local gesture degree-of-freedom parameter sequences; the lip-sync primitives include facial weight temporal sequences corresponding to phonemes; and the interactive micro-response primitives include single-action keyframe units. An animation semantic topology tagging system is established to annotate all primitive units in multiple dimensions.

3. The data asset-driven digital human performance adjustment method for digital product development according to claim 2, characterized in that, The annotation dimensions in the asset pool construction phase include animation parameter type, kinematic attributes, semantic attributes, style attributes, and compatibility attributes; a unified animation asset coding standard is established simultaneously, a unique asset identifier is assigned to each primitive, and layered storage is performed at three levels of granularity: primitive layer, sequence layer, and fragment layer. For each primitive, the coordinates of the parametric manifold space and the interpolation compatibility are calibrated, the fusion boundaries of different primitives in a unified low-dimensional parametric manifold space are defined, and an animation primitive asset pool is constructed.

4. The data asset-driven digital human performance adjustment method for digital product development according to claim 3, characterized in that, In the graph construction phase, typical application scenarios of digital products are sorted out, and core feature vectors of scenarios are extracted, including four dimensions: interaction type, user group attributes, product business tone, and core business objectives. Based on animation psychology and motion biomechanics theory, a quantitative mapping relationship between scenario features and animation parameter space is established, and the constraint boundaries of digital human animation in different scenarios are clarified.

5. The data asset-driven digital human performance adjustment method for digital product development according to claim 4, characterized in that, The constraint boundaries include the facial deformation weight range, the bone joint rotation degree of freedom threshold, the keyframe interpolation curve slope range, the lip-sync amplitude scaling factor, and the motion feedback delay window. The constructed constraint association graph nodes are divided into scene feature nodes, animation parameter nodes, and constraint rule nodes, and the edges between nodes correspond to constraint strength coefficients. Based on the association graph, the primitive scene adaptability score and interpolation weight baseline are calculated for each animation primitive in the asset pool, and the available parameter boundaries under the corresponding scene are marked for each primitive, forming the primitive call baseline configuration for different business scenarios.

6. The data asset-driven digital human performance adjustment method for digital product development according to claim 5, characterized in that, In the model building phase, a baseline model for generating animation parameters driven by multimodal input is constructed. The model is set with three input branches: speech features, text semantic features, and interactive trigger signals. After cross-modal attention fusion and decoding, the input features output three types of animation parameter sequences that can directly drive the rendering engine.

7. The data asset-driven digital human performance adjustment method for digital product development according to claim 6, characterized in that, The three types of animation parameter sequences are the face blendshape weight sequence, the bone rotation keyframe sequence, and the lip-sync parameter sequence; the animation primitive adaptation interface of the feature decoding layer corresponds one-to-one with the three types of parameter branches, and supports the dynamic injection of the feature distribution parameters of the corresponding primitives. The built-in distribution alignment module of the interface can align the original parameter distribution generated by the model to the parameter space of the injected primitives; after completing the general pre-training, the standardized interface is encapsulated, and the model only learns the temporal coherence and multimodal alignment logic of the animation parameters, while the performance style and semantic features are defined by the injected animation primitive assets.

8. A data asset-driven digital human performance adjustment method for digital product development according to claim 7, characterized in that, During the requirement analysis phase, a perception module is deployed on the digital product client to collect user interaction data including voice semantic text, touch operation behavior, and page dwell time, as well as product business status data including the current page node, business process stage, and trigger event type. Feature extraction and normalization are performed on real-time data to generate the current scene requirement vector. After input constraint association graph matching and reasoning, the set of hard constraints generated by animation parameters and the constraint priority weights of the three types of parameter branches are obtained.

9. A data asset-driven digital human performance adjustment method for digital product development according to claim 8, characterized in that, In the parameter generation stage, based on the target parameter constraints, and using semantic matching degree, parameter compatibility degree, and scene adaptability degree as comprehensive retrieval indicators, an approximate nearest neighbor retrieval is performed in the animation primitive asset pool to select the top three candidate primitive units with the highest matching degree. An animation parameter manifold space weighted interpolation algorithm is adopted to map candidate primitives to a unified low-dimensional parameter manifold space. The comprehensive retrieval matching degree weight of each primitive and the target constraint are combined to complete continuous weighted interpolation and generate an animation parameter set. Based on the skeleton binding constraints of the digital human 3D model and the Blendshape deformation boundary, kinematic compliance pre-verification is performed, and parameters that exceed the limit are truncated and smoothed.

10. A data asset-driven digital human performance adjustment method for digital product development according to claim 9, characterized in that, In the output stage, spatiotemporal consistency verification and hierarchical clipping detection are performed on the animation parameter set; in the temporal dimension, cubic Bézier curves are used to achieve inter-frame smoothing, and low-pass filtering is applied to the deformation weights of high-frequency jitter; in the spatial dimension, based on the animation semantic alignment rules, the semantic strength consistency of the three types of parameters is verified, and amplitude scaling correction is applied to deviations exceeding the threshold; collision risks are verified sequentially from the skeleton and mesh levels, and the parameters of risk frames are fine-tuned; the verified parameter sequence is output to the real-time rendering engine, while providing downsampling and precision simplification options for multi-terminal adaptation; During the optimization phase, two multi-dimensional indicators, namely animation technical quality and business effect, are collected. A comprehensive quality assessment model is used to complete the quantitative scoring, dynamically adjust the primitive adaptation and interpolation weight, extract high-quality animation sequences as standard primitive assets for storage, and regularly perform clustering deduplication, quality classification, and cleanup of low-quality and low-reusability assets in the asset pool.

Images