Action migration method and device, electronic equipment and computer storage medium

Abstract

The embodiment of the invention provides an action migration method and device, electronic equipment and a computer storage medium. The action migration method comprises the following steps: displaying a target role model; acquiring a reference action video; the reference action video comprises a reference role for executing the reference action; and in response to a preset trigger operation, according to a preset execution sequence among a plurality of action migration sub-steps, calling the artificial intelligence models corresponding to the action migration sub-steps in sequence, so as to execute the corresponding action migration sub-steps through the artificial intelligence models, and migrating the reference action executed by the reference role to the target role model. Through the embodiment of the invention, an end-to-end automatic process from target role model display to target role model driving is realized based on the artificial intelligence model, the implementation threshold of action migration is reduced, and the efficiency and generalization ability of action migration are improved while the implementation cost is saved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a motion transfer method, apparatus, electronic device, and computer storage medium. Background Technology

[0002] Cross-character motion transfer refers to the process of completely transferring the motion of an existing character (reference character or source character) to another character model (target character model), thereby enabling the target character model to visually reproduce the "same motion". For example, cross-character motion transfer is often involved in the production of 3D character animation, allowing the 3D character model to perform specified actions.

[0003] Currently, the main method for migrating virtual characters' movements is through manual processing, which is time-consuming, costly, and unfriendly to non-professional users. Summary of the Invention

[0004] In view of this, embodiments of this application provide an action transfer scheme to at least partially solve the above-mentioned problems.

[0005] According to a first aspect of the embodiments of this application, an action transfer method is provided, including: Display the target character model; Obtain a reference action video; the reference action video includes a reference character performing the reference action; In response to a preset trigger operation, the AI ​​model corresponding to each action transfer sub-step is called sequentially according to the execution order of multiple action transfer sub-steps set in advance, so that each AI model executes the corresponding action transfer sub-step and transfers the reference action performed by the reference role to the target role model.

[0006] According to a second aspect of the embodiments of this application, an action transfer device is provided, comprising: The display module is used to display the target character model; The video acquisition module is used to acquire a reference action video; the reference action video includes a reference character performing the reference action; The action transfer module is used to respond to a preset trigger operation and sequentially call the artificial intelligence model corresponding to each action transfer sub-step according to the execution order of multiple preset action transfer sub-steps, so that each artificial intelligence model executes the corresponding action transfer sub-step and transfers the reference action performed by the reference character to the target character model.

[0007] According to a third aspect of the present application, an electronic device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; the memory is used to store at least one executable instruction, wherein the executable instruction causes the processor to perform an operation corresponding to the method described in the first aspect.

[0008] According to a fourth aspect of the embodiments of this application, a computer storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0009] According to a fifth aspect of the embodiments of this application, a computer program product is provided, which includes computer instructions that instruct a computing device to perform an operation corresponding to the method described in the first aspect.

[0010] According to the action transfer scheme provided in the embodiments of this application, a target character model can be displayed; a reference action video can be obtained; the reference action video includes a reference character performing the reference action; in response to a preset trigger operation, according to the execution order between multiple preset action transfer sub-steps, the artificial intelligence model corresponding to each action transfer sub-step is called in sequence, so that each artificial intelligence model executes the corresponding action transfer sub-step respectively, and the reference action performed by the reference character is transferred to the target character model.

[0011] This application embodiment integrates a streamlined processing platform for automatically implementing action transfer according to the execution order of the various action transfer sub-steps involved in the action transfer task. Furthermore, a corresponding artificial intelligence model is set for each action transfer sub-step. Therefore, through this application embodiment, an end-to-end automated process from target role model display to target role model-driven execution is realized based on the artificial intelligence model, lowering the implementation threshold of action transfer, saving implementation costs, and improving the efficiency and generalization ability of action transfer. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0013] Figure 1 This is a flowchart illustrating the steps of an action transfer method according to an embodiment of this application. Figure 2 This is a schematic diagram illustrating an interface according to an embodiment of this application; Figure 3 This is a schematic diagram illustrating another interface according to an embodiment of this application; Figure 4 This is a schematic diagram illustrating another interface according to an embodiment of this application; Figure 5 This is a schematic diagram illustrating another interface according to an embodiment of this application; Figure 6 This is a schematic diagram illustrating another interface according to an embodiment of this application; Figure 7 This is a structural block diagram of a motion transfer device according to an embodiment of this application; Figure 8 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in the embodiments of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art should fall within the protection scope of the embodiments of this application.

[0015] Reference Figure 1 , Figure 1 This is a flowchart illustrating the steps of an action migration method according to an embodiment of this application. The action migration method provided in this embodiment can be executed by any suitable electronic device with data processing capabilities. For example, the electronic device can be a server, a PC, etc. The action migration method may include the following steps: Step 102: Display the target character model.

[0016] Schematic illustration: The target character model shown in this step can be a pre-created digital model, such as a two-dimensional or three-dimensional character model. This application embodiment does not limit the spatial topology or construction method of the target character model. For example, regarding spatial topology, the target character model can be a human model, an animal model, etc. Regarding the model construction method, it can be manually constructed by performing a series of interactive operations within the interface provided by the modeling software, or it can be automatically output using artificial intelligence technology and a text-based neural network model.

[0017] Step 104: Obtain the reference action video; the reference action video contains a reference character performing the reference action.

[0018] Indicatively, the reference action video contains a reference character. As the video timestamp is updated, the actions performed by the reference character can change dynamically, thus forming a series of reference action sequences. The aforementioned reference action sequence corresponds to a set of dynamic reference actions.

[0019] In this embodiment, the type of reference character is not limited. For example, it can be a virtual animated character, a real person, or an animal. Similarly, the reference action in this embodiment can be any action, such as raising both hands above the head, jumping with both feet, jumping with one foot, shooting a basketball with a bent knee, etc.

[0020] In this embodiment of the application, the execution order of the above steps 102 and 104 is not limited. For example, step 102 can be executed first and then step 104 can be executed; step 104 can be executed first and then step 102 can be executed; or steps 102 and 104 can be executed in parallel.

[0021] Step 106: In response to the preset trigger operation, according to the execution order of multiple preset action transfer sub-steps, the artificial intelligence model corresponding to each action transfer sub-step is called in sequence, so that each artificial intelligence model executes the corresponding action transfer sub-step and transfers the reference action performed by the reference character to the target character model.

[0022] Schematic illustration: The action transfer step can be pre-divided into multiple sequentially executed action transfer sub-steps, and a corresponding artificial intelligence model can be deployed for each action transfer sub-step, thereby leveraging the potential of the artificial intelligence model to simplify and accelerate the execution process of each sub-step. In this embodiment, there are no limitations on the specific model type of the artificial intelligence model or the specific values ​​of the model parameters; any suitable artificial intelligence model can be used, such as a pre-trained neural network model.

[0023] In this way, once the user's preset trigger operation is received, the corresponding artificial intelligence model can be called sequentially to execute the corresponding action transfer sub-steps according to the execution order between each action transfer sub-step, until all action transfer sub-steps have been executed, and the reference action performed by the reference character in the reference action video is transferred to the target character model.

[0024] Schematically, the motion transfer process can be broken down into the following sequentially executed sub-steps: skeleton rigging, skinning weight optimization, and motion redirection. The skeleton rigging sub-step refers to creating the skeletal structure of the target character model and associating this structure with each vertex in the model (i.e., assigning skinning weight values ​​to each vertex, which represent the degree of deformation of the vertex as it moves with the bones). Through the skeleton rigging sub-step, the movement of each vertex in the target character model can be easily controlled by moving the bones, without manually adjusting the position of the vertices. The skinning weight optimization sub-step refers to optimizing and correcting the skinning weight values ​​of each vertex in the model (or parts of the model) determined in the skeleton rigging sub-step, thereby making the subsequent movements of the target character model more natural and fluid. The action redirection sub-step can refer to rearranging the reference action sequence performed by the reference character into a new action sequence suitable for the target character model. Furthermore, the above rearrangement process will maintain the semantics of the action sequence. For example, if the reference character performs a walking action, the target character model will also perform a walking action; if the reference character performs a waving action, the target character model will also perform a waving action.

[0025] This application embodiment integrates a streamlined processing platform for automatically implementing action migration according to the execution order of the various action migration sub-steps involved in the action migration process. Furthermore, a corresponding artificial intelligence model is set for each action migration sub-step. Therefore, through this application embodiment, an end-to-end automated process from target role model display to target role model-driven execution is realized based on the artificial intelligence model, lowering the implementation threshold of action migration, saving implementation costs, and improving the efficiency and generalization ability of action migration.

[0026] Optionally, in some embodiments, after displaying the target character model, the above-described motion transfer method may further include: Display action transition process information; action transition process information includes: the name information of each action transition sub-step, and the execution order information between each action transition sub-step.

[0027] Schematic, see Figure 2 , Figure 2 This is a schematic diagram illustrating an interface according to an embodiment of this application. Figure 2 The interface can simultaneously display the target character model and motion transfer process information. Indicatively, this motion transfer process information represents the following: the motion transfer process involves three sub-steps: skeleton rigging, skinning weight optimization, and motion retargeting. The execution order of these three sub-steps is as follows: first, the skeleton rigging sub-step; then, the skinning weight optimization sub-step; and finally, the motion retargeting sub-step.

[0028] In the embodiments described above, before initiating the action migration process, the user is shown what the various action migration sub-steps are about to be executed, and the execution order of these sub-steps. By displaying this action migration process information, users can better understand the specific implementation process of the action migration, build the overall architecture of the action migration process, and understand the logic and dependencies between the various sub-steps, thereby improving the transparency of the action migration implementation process and enhancing the user experience.

[0029] Optionally, in some embodiments, after displaying the target character model, the motion transfer method may further include: Displays introductory information for the currently pending sub-step; the currently pending sub-step is the earliest execution sub-step among the unexecuted action transition sub-steps.

[0030] The process described above, which responds to a preset trigger operation and sequentially calls the AI ​​model corresponding to each action transition sub-step according to the pre-defined execution order, may include: In response to a trigger operation on the current sub-step to be executed, the target artificial intelligence model corresponding to the current sub-step to be executed is invoked, and the current sub-step to be executed is executed through the target artificial intelligence model to obtain the execution result; According to the pre-defined execution order of multiple action migration sub-steps, the current sub-step to be executed is redefined, and the step that displays the introductory information for the current sub-step to be executed is returned, until the reference action performed by the reference character is migrated to the target character model.

[0031] Schematic illustration: In the specific implementation of the above embodiments of this application, for multiple action migration sub-steps, the currently pending sub-step, which has not yet been executed at the current moment, can be determined according to the execution order. Then, the user is shown an introduction to this currently pending sub-step. When a user triggers an operation on this currently pending sub-step, the introduction to this currently pending sub-step is executed. Afterwards, a new currently pending sub-step is determined according to the execution order, and the user is shown an introduction to this new currently pending sub-step. When a user triggers an operation on this new currently pending sub-step, the new currently pending sub-step is executed. This process is repeated until the last action migration sub-step is executed, thus migrating the reference action performed by the reference character to the target character model.

[0032] For example, at the start of the motion transfer process, the current sub-step to be executed can be identified as the skeleton binding sub-step. At this point, information about the skeleton binding sub-step can be displayed to the user, allowing them to understand the execution process in more detail. Upon receiving a user's trigger operation for the skeleton binding sub-step, the corresponding AI model is invoked to execute the skeleton binding sub-step, thus obtaining the skeleton binding result. After the skeleton binding sub-step is completed, the skinning weight optimization sub-step can be identified as the new current sub-step to be executed. Therefore, information about the skinning weight optimization sub-step can be displayed to the user. Upon receiving a user's trigger operation for the skinning weight optimization sub-step, the corresponding AI model is invoked to execute the skinning weight optimization sub-step, thus obtaining the optimization result. After the skinning weight optimization sub-step is completed, the motion redirection sub-step can be identified as the new current current sub-step to be executed. Therefore, information about the motion redirection sub-step can be displayed to the user. Upon receiving a user's trigger operation for the motion redirection sub-step, the corresponding AI model is invoked to execute the motion redirection sub-step, thus obtaining the final motion transfer result, that is, transferring the reference action executed by the reference character to the target character model.

[0033] In this embodiment of the application, the specific content of the introductory information for the currently to be executed sub-step is not limited. For example, the introductory information may be used to describe at least one of the following aspects of the currently to be executed sub-step: the name of the currently to be executed sub-step, the function of the currently to be executed sub-step, the principle of the currently to be executed sub-step, and the processing flow of the currently to be executed sub-step.

[0034] Furthermore, in this embodiment, the specific operation content of the triggering operation for the currently executed sub-step is not limited, and can be customized according to the actual situation. For example, the triggering operation can be a click operation on a physical component (such as a mouse or keyboard light) of the electronic device executing the method of this embodiment, or it can be a triggering operation (such as a single click, double click, drag, etc.) on a preset component displayed on the display interface, and so on.

[0035] In the embodiments described above, during the execution of the action migration process, the currently pending sub-step is dynamically determined, and introductory information about that sub-step is displayed to the user. Furthermore, the execution of the currently pending sub-step is triggered based on the user's action on it. Through this method, the user can decide whether and when to execute the sub-step based on the provided information. Therefore, this enhances the user's sense of participation and control, thereby further improving the user experience.

[0036] Optionally, in some embodiments, for the skeletal binding sub-step, the process of invoking the artificial intelligence model corresponding to the action transfer sub-step to execute the corresponding action transfer sub-step through the artificial intelligence model may include: Obtain semantic segmentation results for the target role model; The model data and semantic segmentation results corresponding to the target character model are input into the skeleton recognition model, and the skeleton recognition model outputs the skeleton recognition results; the skeleton recognition results include: joint positions and joint names; Based on the joint positions and joint names, determine the bone positions and bone names of the target character model; Based on the bone positions and bone names of the target character model, determine the bone hierarchy of the target character model; Based on the bone position, bone name, and bone hierarchy of the target character model, the skinning weights of each vertex in the target character model are initialized.

[0037] Schematic, the AI ​​model corresponding to the skeleton binding sub-step can include multiple parts for implementing different functions: a skeleton recognition model, a skeleton placement layer, a hierarchy construction layer, and an initialization layer. Specifically, the skeleton recognition model performs skeleton recognition on the target character model, outputting the joint names and positions of the joints contained in the target character model; the skeleton placement layer determines the bone names and positions of each bone in the target character model based on the joint positions and names output by the skeleton recognition; the hierarchy construction layer determines the hierarchical relationship between the bones based on the bone names and positions determined by the skeleton placement layer; and the initialization layer initializes the skinning weight values ​​of each vertex in the target character model based on the bone positions and names determined by the skeleton placement layer and the hierarchical relationship determined by the hierarchy construction layer.

[0038] The implementation methods for each part are explained below: Regarding skeleton recognition models, the joint names and positions of the joints contained in the target character model can be predicted based on the model data and semantic segmentation results. The semantic segmentation result of the target character model refers to the result obtained by semantically segmenting the target character model. Semantic segmentation refers to dividing each model unit (such as a vertex, model facet, or model texel) in the target character model into different semantic regions. These semantic regions are divided according to the model parts of the target character model, with different semantic regions corresponding to different model parts. Furthermore, for easier differentiation, corresponding semantic labels can be set for each semantic region. For example, the semantic label "hand" can be set for the hand semantic region, and the semantic label "head" can be set for the head semantic region. The semantic segmentation result includes the correspondence (mapping relationship) between each model unit and the semantic label. For example, when the smallest model unit in the target character model is a model vertex, the above semantic segmentation result may include: model vertex labeled 1 corresponds to the semantic label "hand," model vertex labeled 2 corresponds to the semantic label "head," and so on.

[0039] For example, semantic segmentation results for a target character model can be obtained as follows: The target character model is input into a pre-trained semantic segmentation model, which outputs the semantic segmentation results for the target character model, thus splitting the target character model into different model parts. For instance, taking a 3D animated character as an example, the 3D semantic segmentation model can split the 3D animated character into model parts such as head, hair, legs, and feet. Furthermore, for easier differentiation, a 3D masking mechanism can be used, where different mask values ​​(semantic labels) can be used to represent different model parts.

[0040] After obtaining the semantic segmentation results for the target character model, a skeleton recognition model can be invoked. The semantic segmentation results and the corresponding model data of the target character model are input into the skeleton recognition model. The skeleton recognition model then predicts the joint positions and names of each joint in the target character model. This process can be called skeleton recognition. For example, the skeleton recognition model can be a recurrent neural network model, a convolutional neural network model, a Transformer model, etc.

[0041] Furthermore, biological prior knowledge graphs can be incorporated into the skeleton recognition process. Each node in the biological prior knowledge graph corresponds to a model part of the target character model. The node structure stores the number of joints that should appear in the corresponding model part, as well as the area within which each joint should appear. During skeleton recognition, in addition to inputting the semantic segmentation results and corresponding model data of the target character model into the skeleton recognition model, the aforementioned biological prior knowledge graph can also be input. This allows the biological prior knowledge graph to guide the skeleton recognition model in predicting the joint positions and names within the areas where joints should appear. This approach improves the efficiency of skeleton recognition and makes the recognized results more reasonable.

[0042] After completing the skeleton recognition steps described above, the skeleton placement layer, hierarchy construction layer, and initialization layer can be used sequentially to perform skeleton placement, hierarchy construction, and weight initialization to complete the skeleton binding sub-step. Skeleton placement refers to determining the skeleton positions and names of the target character model based on the predicted joint positions and names. Skeleton placement can be implemented by: determining the parent joint of each joint based on its name to form parent-child joint pairs; for each parent-child joint pair, forming a directed line segment with the parent node as the starting point and the child joint as the ending point; the starting point, ending point, and direction vector of this line segment together constitute the skeleton position; furthermore, the skeleton name can be determined based on the aforementioned skeleton position. Hierarchical construction refers to determining the hierarchical relationship between the skeletons of the target character model (e.g., thigh-lower leg-foot) based on the skeleton positions and names of the target character model. The implementation of hierarchical construction can include: using each bone obtained through the bone placement step as directed edges to construct a directed graph, where the nodes in the graph correspond to the joints of the target character model; determining the root node in the directed graph, and performing a depth-first search starting from the root node, for example: depth 0 corresponds to the root node; depth 1 corresponds to the child node of the root node; depth 2 corresponds to the child node of the root node, and so on, until the depth value of each node is obtained, which is the hierarchy of the node (i.e., the joint).

[0043] Initializing weights refers to setting initial skinning weight values ​​for each vertex in the target character model based on the bone position, bone name, and bone hierarchy. The implementation of weight initialization can include: for each vertex, traversing each bone and setting an initial skinning weight value for each bone. The setting principle can include: calculating the distance between the vertex and the bone; the closer the distance, the larger the weight value; when the distance exceeds a preset distance threshold, the weight value is set to 0.

[0044] In this embodiment, for the artificial intelligence model corresponding to the skeleton binding sub-step, during the model training phase, it is usually sufficient to adjust the parameters of the skeleton recognition model, without needing to adjust the parameters of other parts (skeleton placement layer, hierarchical construction layer, and initialization layer). The training process of the skeleton recognition model may include: obtaining model data and semantic segmentation results corresponding to the digital character model as training samples; inputting the above training samples into the skeleton recognition model; outputting prediction results through the skeleton recognition model: joint positions and joint names; calculating loss values ​​based on the prediction results and training labels (the names and positions of each joint actually included in the above digital character model); adjusting model parameters according to the loss values ​​to obtain the trained skeleton recognition model.

[0045] In the embodiments described above, an artificial intelligence model is embedded during the execution of the skeletal rigging sub-step. This model simplifies and accelerates the skeletal rigging process, eliminating the need for manual intervention by the user. Instead, the user can seamlessly switch to the AI-driven intelligent skeletal rigging mode through a simple triggering operation. Therefore, the embodiments of this application improve the efficiency and quality of skeletal rigging.

[0046] Furthermore, in the above embodiments, before performing skeleton recognition based on the skeletal recognition model, the semantic segmentation result of the target character model is first obtained. Then, the semantic segmentation result and the model data are simultaneously input into the skeleton recognition model. The semantic segmentation result can provide the feature extractor (such as PointNet+) in the skeleton recognition model with prior knowledge of "which model part each vertex belongs to." That is, through the semantic segmentation result, the skeletal recognition model can clearly identify the semantic region to which each vertex belongs. Subsequently, joint coordinate regression within each semantic region can generate skeleton recognition results that conform to the internal structure of each semantic region. Therefore, through the above-mentioned scheme provided by the embodiments of this application, accurate skeleton recognition of target character models with arbitrary topological structures or arbitrary proportions that are not humanoid can be achieved, further improving the quality of skeleton binding.

[0047] See Figure 3 , Figure 3 This is a schematic diagram illustrating another interface according to an embodiment of this application. Figure 3 The image shown is a schematic diagram of the interface when the current sub-step to be executed is a skeleton binding sub-step. Figure 3 and Figure 2 What they have in common is that the display interface shows the target character model and motion transfer process information. The difference is... Figure 3The process information highlights the skeleton binding sub-step and displays introductory information for the skeleton binding sub-step (i.e., the currently pending sub-step). For example, the introductory information includes: an overall overview of the skeleton binding sub-step, "We will automatically identify character skeletons and create a skeleton hierarchy," and the internal processing flow of the skeleton binding sub-step: "Skeleton identification, skeleton placement, hierarchy construction, and weight initialization."

[0048] in addition, Figure 3 The interface shown also displays trigger controls for the skeleton binding sub-step, that is... Figure 3 The virtual icon in the middle reads "Start Skeletal Binding". When a user wants to initiate the skeletal binding sub-step, they can trigger (e.g., by clicking) the virtual icon. Afterward, the skeletal binding sub-step will be executed automatically until the result is obtained.

[0049] Optionally, in some embodiments, for the skin weight optimization sub-step, calling the artificial intelligence model corresponding to the action transfer sub-step, so as to execute the corresponding action transfer sub-step through the artificial intelligence model, may include: Displays icons representing the different parts of the target character model; In response to the selection operation of the target model part icon, obtain the target part data corresponding to the target model part and the bone position and bone name of each target bone contained in the target model part; Input the target part data, the bone position of each target bone, and the bone name into the skin weight optimization model. The optimized skin weight of each vertex in the target model part is then predicted by the skin weight optimization model.

[0050] Indicatively, in this embodiment, the artificial intelligence model corresponding to the skin weight optimization sub-step can be a skin weight optimization model. The specific model structure and parameter settings of the skin weight optimization model are not limited; they can be customized or selected according to actual conditions. For example, a graph attention network can automatically assign higher weight values ​​to joints that have a significant impact on the model vertices, and assign smaller weight values ​​to joints that have no or minimal impact on the model vertices, thus making the final skin weight values ​​transition more naturally without subsequent manual smoothing adjustments. Therefore, a graph attention network can be used as the skin weight optimization model. The training process of the skin weight optimization model can include: obtaining model data corresponding to the model parts, the bone positions and bone names of each bone in the model parts as training samples, inputting the above training samples into the skin weight optimization model, and outputting prediction results through the skin weight optimization model: the skin weight values ​​of each vertex in the model parts; calculating the loss value based on the prediction results and training labels (the true values ​​of the skin weights of each vertex in the model parts, which can be manually adjusted by the user in the content creation software), adjusting the model parameters according to the loss value, and obtaining the trained skin weight optimization model.

[0051] In the embodiments described above, before executing the skin weight optimization sub-step, icons representing different model parts within the target character model can be displayed. Users can select the target model part to be optimized from multiple model parts according to their needs. Then, the optimized skin weights of each vertex in the target model part can be re-predicted by calling the skin weight optimization model. Alternatively, users can also treat the complete target character model as a whole and optimize the skin weight values ​​for the entire model. The specific optimization process is similar to the optimization process for the target model part described above, and will not be repeated here.

[0052] See Figure 4 , Figure 4 This is a schematic diagram of the interface before the skinning weight optimization sub-step is executed. Figure 4 The left side displays the target character model, while the right side, in addition to showing the motion transfer process information, also displays information about the skin weight optimization sub-step: "Overall AI skin weight optimization, or, select individual part optimization." Additionally, it displays icons corresponding to different parts of the target character model, such as the icon for the "head," the icon for the "torso," and so on.

[0053] in addition, Figure 4 The document also showcases trigger controls for the skinning weight optimization sub-step, namely... Figure 4The virtual icon in the middle reads "Start Weight Optimization". Users can first select the target model part icon, and then trigger the virtual icon. After that, the skinning weight optimization sub-step will be automatically executed for the target model part corresponding to the target model part icon until the execution result is obtained.

[0054] In this embodiment, a skin weight optimization model is embedded during the skin weight optimization sub-step. This model simplifies and accelerates the weight optimization process, eliminating the need for manual intervention by the user. Instead, the user can seamlessly switch to the intelligent weight optimization mode driven by the neural network model through a simple triggering operation. Therefore, this embodiment improves the efficiency and quality of skin weight optimization.

[0055] Furthermore, through the above process, users can selectively perform weight optimization operations: they can optimize the weights of the entire target character model, or they can optimize specific parts of the model. This improves the convenience and flexibility of weight optimization.

[0056] Optionally, in some embodiments, the process of calling the artificial intelligence model corresponding to the action transfer sub-step for the action redirection sub-step, so as to execute the corresponding action transfer sub-step through the artificial intelligence model, may include: The reference motion video is parsed to obtain the reference skeleton topology information of the reference character, as well as the reference skeleton position information corresponding to each reference motion video frame. The reference skeleton topology information includes: the skeleton name of the reference character and the skeleton hierarchy relationship; the reference skeleton position information includes the skeleton position of the reference character. Input the reference skeleton topology information and the target skeleton topology information of the target character model into the skeleton mapping model, perform semantic alignment operation through the skeleton mapping model, and output the skeleton mapping relationship between the reference character and the target character model. The reference bone position information, bone mapping relationship and bone position of the target character model corresponding to each reference action video frame are input into the motion retargeting model. The motion retargeting model outputs the rotation angle sequence of each bone in the target character model. Based on the rotation angle sequence and the skinning weight values ​​of each vertex in the target character model, the positions of each bone and vertex in the target character model are dynamically adjusted to transfer the reference actions performed by the reference character to the target character model.

[0057] Indicatively, regarding the motion retargeting sub-step, after acquiring the reference motion video, the video can be parsed to obtain the reference skeleton topology information (including bone names and bone hierarchy relationships) and bone positions of the reference character. For example, keypoint (such as joints) sequences of the reference character can be extracted from the reference video frames; then, 3D upscaling is performed based on the above keypoint sequences to obtain the 3D reference skeleton topology information and bone positions of the reference character. That is, 2D keypoints are extracted frame by frame from the reference motion video. For example, 2D keypoints can be extracted from each reference motion video frame using a 2D human pose estimation model such as ViTPose to obtain a 2D keypoint sequence; then, 3D upscaling is performed on the above 2D keypoint sequence (using camera calibration algorithms or pre-trained 3D upscaling network models such as VideoPose3D, LiftFormer, PoseNet-3D, etc.) to obtain a 3D skeleton motion sequence and the bone names, bone positions, and bone hierarchy relationships of each bone of the reference character.

[0058] The AI ​​model corresponding to the motion retargeting sub-step can include a skeleton mapping model and a motion retargeting model. After obtaining the reference skeleton topology information and bone positions of the reference character, the skeleton mapping model can be called first to perform semantic alignment based on the reference skeleton topology information of the reference character and the target skeleton topology information of the target character model, thereby predicting the skeleton mapping relationship between the reference character and the target character model, that is, which bone in the reference character corresponds to which bone in the target character model. Then, the motion retargeting model is called to redirect the reference bone positions in each motion reference video frame to the corresponding bones in the target character model, thereby obtaining the rotation angle sequence of each bone in the target character model, that is, the rotation angle data of each bone in the target character model in each video frame (usually identified by quaternions or Euler angles).

[0059] The aforementioned action redirection model can be based on an RTN (Recursive Transition Network), such as a recurrent neural network (RNN), convolutional neural network (CNN), or Transformer model, etc. Similarly, the aforementioned skeleton mapping model can also be a RNN, CNN, or Transformer model, etc.

[0060] In this embodiment of the application, during the model training phase, the skeletal mapping model and the action redirection model can be trained independently, or they can be trained jointly.

[0061] For independent training, the training process of the skeletal mapping model may include: acquiring reference skeletal topology information and the skeletal topology information of the digital character model as training samples; inputting the training samples into the skeletal mapping model; outputting prediction information through the skeletal mapping model: the predicted skeletal mapping relationship between the reference character and the target character model; calculating the loss value based on the prediction information and training labels (the actual skeletal mapping relationship between the reference character and the target character model); adjusting the model parameters based on the loss value to obtain the trained skeletal mapping model. The training process of the motion retargeting model may include: acquiring reference skeletal position information, skeletal mapping relationship, and skeletal positions of the digital character model corresponding to reference motion video frames as training samples; inputting the training samples into the motion retargeting model; outputting prediction information through the motion retargeting model: the predicted rotation angle sequence of each bone in the digital character model; calculating the loss value based on the prediction information and training labels (the actual rotation angle sequence of each bone in the digital character model, which can be manually adjusted in content creation software); adjusting the model parameters based on the loss value to obtain the trained motion retargeting model.

[0062] Regarding joint training, the training process may include: acquiring reference skeletal topology information and skeletal topology information of the digital character model as training samples; inputting the above training samples into the artificial intelligence model corresponding to the action redirection sub-step, which includes a skeletal mapping model and an action redirection model; outputting prediction information through the above artificial intelligence model: the predicted rotation angle sequence of each bone in the digital character model; calculating the loss value based on the prediction information and training labels (the actual rotation angle sequence of each bone in the digital character model, which can be manually adjusted by humans in content creation software); adjusting the parameters of the skeletal mapping model and the action redirection model according to the loss value to obtain the trained artificial intelligence model.

[0063] After obtaining the above rotation angle sequence through the artificial intelligence model corresponding to the action redirection sub-step, the position of each bone in the target character model can be dynamically adjusted according to the rotation angle sequence. The position of each vertex can be adjusted according to the skinning weight value of each vertex in the target character model (this value can be the weight value obtained by executing the skinning weight optimization sub-step), thereby transferring the reference action performed by the reference character to the target character model.

[0064] In the embodiments described above, a skeleton mapping model and an action redirection model are embedded during the execution of the action redirection sub-step. These two models simplify and accelerate the weight optimization process, eliminating the need for manual processing by the user. Instead, the user can seamlessly switch to the intelligent action redirection mode driven by the artificial intelligence model through a simple triggering operation. Therefore, the embodiments of this application improve the efficiency and quality of action redirection.

[0065] Furthermore, in the embodiments described above, the skeletal mapping model outputs skeletal mapping relationships by performing semantic alignment processing on the bone names and bone hierarchy information of the reference character and the target character model. This method automatically matches semantically similar (functionally similar) bones between the reference and target character models; for example, it can automatically establish a mapping relationship between the "wings" in the reference character and the "arms" in the target character model. Therefore, it supports motion transfer between characters with different skeletal structures, improving the universality of the motion transfer solution.

[0066] Optionally, in some embodiments, after outputting the rotation angle sequence of each bone in the target character model through the motion redirection model, the following post-processing can be performed to correct the rotation angle sequence. The post-processing may include: inverse kinematic end position correction and joint limitation.

[0067] The inverse kinematics end-effector position correction process can include: applying hard constraints to each end-effector joint in the target character model, i.e., setting and locking the end-effector joint positions in preset target positions (e.g., the "toe" joint should contact the ground, the "finger" joint should contact the tabletop, etc.); using an IK (Inverse Kinematics) solver to perform end-effector reverse engineering, recalculating the rotation angle sequence of each bone in the target character model, so that the recalculated rotation angle sequence of each bone can ensure that the end-effector joints accurately reach the target position, avoiding suspension or stretching through the clipping. During the end-effector reverse engineering process using the IK solver, the length of the bones can be compensated (stretched) or compressed to better ensure that the end-effector joints accurately reach the target position. This correction process makes the reference movements performed by the target character model more natural and fluid.

[0068] The joint limiting process can include: in the above IK solver, adding rotation angle restrictions for each joint (such as the rotation angle of the elbow joint being 0-150°, the rotation angle of the knee joint being 0-100°, etc.). By adding the above restrictions, it is possible to effectively prevent problems such as motion distortion, clipping, and abnormal stretching caused by joint reflex, and further improve the naturalness and smoothness of the reference motion.

[0069] Optionally, in some embodiments, the process of acquiring the reference motion video described above may include: Display action icons corresponding to different preset actions; in response to the selection operation of the target action icon, obtain the action video corresponding to the target action icon as a reference action video.

[0070] Indicatively, in this embodiment, the timing of acquiring the reference motion video is not limited and can be any time before the execution of the aforementioned motion redirection sub-step. Considering that the execution of the other sub-steps besides the aforementioned motion redirection sub-step does not require the reference motion video, as an example, the reference motion video can be acquired after the other sub-steps have been completed and before the aforementioned motion redirection sub-step begins execution.

[0071] See Figure 5 , Figure 5 This is a schematic diagram showing the interface before the action redirection sub-step is executed. Figure 5 The left side displays the target character model, and the right side, in addition to displaying the motion transfer process information, also displays an "motion library," which shows multiple motion icons corresponding to different preset actions.

[0072] in addition, Figure 5 The document also showcases trigger controls for action redirection sub-steps, namely... Figure 5 The virtual icon is labeled "Application Action". Users can first select the target action icon, and then trigger the virtual icon to automatically execute the action redirection sub-steps until the result is obtained.

[0073] In the above embodiments of this application, an action library is provided, allowing users to select suitable action videos as reference videos by choosing icons according to their needs. The process is convenient, as users can easily obtain reference action videos through simple icon selection.

[0074] Optionally, in some embodiments, the process of acquiring the reference motion video described above may further include: Displays video upload guidance information; the video upload guidance information is used to guide the execution of the reference action video upload operation; receives the uploaded reference action video.

[0075] Schematic illustration: Similar to the timing of acquiring the reference action video in the above embodiments, the timing of acquiring the reference action video in this application embodiment is not limited and can be any time before the execution of the above-mentioned action redirection sub-step. As an example, the reference action video can also be acquired after other sub-steps have been completed and before the above-mentioned action redirection sub-step begins execution.

[0076] See Figure 6 , Figure 6 This is another schematic diagram showing the interface before the action redirection sub-step is executed. Figure 6 The left side displays the target character model, while the right side, in addition to showing the motion transition process information, also displays the prompt "Upload Driver Video," and includes a video upload control labeled "Click to Upload Video." Furthermore, Figure 6 The document also displays a trigger control for action redirection sub-steps, namely the virtual icon for "Apply Action".

[0077] Users can trigger the video upload control to complete the video upload operation. Afterwards, triggering the virtual icon will automatically execute the action redirection sub-step until the result is obtained.

[0078] In the above embodiments of this application, users can upload action videos that meet their own needs as reference action videos. Therefore, the target character model can perform actions that meet their own needs, thereby better satisfying the user's personalized needs.

[0079] In addition, in some other embodiments of this application, the user may be provided with both of the above-mentioned methods for obtaining reference motion videos at the same time, so that the user can customize and choose the appropriate method according to their own situation.

[0080] See Figure 7 , Figure 7 This is a structural block diagram of a motion transfer device according to an embodiment of this application. The motion transfer device provided in this application includes: Display module 702 is used to display the target character model; The video acquisition module 704 is used to acquire a reference action video; the reference action video contains a reference character performing the reference action; The action transfer module 706 is used to respond to a preset trigger operation and sequentially call the artificial intelligence model corresponding to each action transfer sub-step according to the execution order of multiple preset action transfer sub-steps, so that each artificial intelligence model executes the corresponding action transfer sub-step and transfers the reference action performed by the reference character to the target character model.

[0081] Optionally, in some embodiments, after displaying the target character model, the display module 702 is further configured to: Display action transition process information; action transition process information includes: the name information of each action transition sub-step, and the execution order information between each action transition sub-step.

[0082] Optionally, in some embodiments, after displaying the target character model, the display module 702 is further configured to: Displays introductory information for the currently pending sub-step; the currently pending sub-step is the earliest executed sub-step among the unexecuted sub-steps. Action transition module 706 is specifically used for: In response to a trigger operation on the current sub-step to be executed, the target artificial intelligence model corresponding to the current sub-step to be executed is invoked, and the current sub-step to be executed is executed through the target artificial intelligence model to obtain the execution result; According to the pre-defined execution order of multiple action migration sub-steps, the current sub-step to be executed is redefined, and the step that displays the introductory information for the current sub-step to be executed is returned, until the reference action performed by the reference character is migrated to the target character model.

[0083] Optionally, in some embodiments, the multiple motion transfer sub-steps sequentially include: a skeleton binding sub-step, a skinning weight optimization sub-step, and a motion redirection sub-step.

[0084] Optionally, in some embodiments, the motion transfer module 706, during the process of executing the invocation of the artificial intelligence model corresponding to the motion transfer sub-step for the skeleton binding sub-step, so as to execute the corresponding motion transfer sub-step through the artificial intelligence model, is specifically used for: Obtain semantic segmentation results for the target role model; The model data and semantic segmentation results corresponding to the target character model are input into the skeleton recognition model, and the skeleton recognition model outputs the skeleton recognition results; the skeleton recognition results include: joint positions and joint names; Based on the joint positions and joint names, determine the bone positions and bone names of the target character model; Based on the bone positions and bone names of the target character model, determine the bone hierarchy of the target character model; Based on the bone position, bone name, and bone hierarchy of the target character model, the skinning weights of each vertex in the target character model are initialized.

[0085] Optionally, in some embodiments, the action transfer module 706, during the process of executing the invocation of the artificial intelligence model corresponding to the action transfer sub-step for the skin weight optimization sub-step, so as to execute the corresponding action transfer sub-step through the artificial intelligence model, is specifically used for: Displays icons representing the different parts of the target character model; In response to the selection operation of the target model part icon, obtain the target part data corresponding to the target model part and the bone position and bone name of each target bone contained in the target model part; Input the target part data, the bone position of each target bone, and the bone name into the skin weight optimization model. The optimized skin weight of each vertex in the target model part is then predicted by the skin weight optimization model.

[0086] Optionally, in some embodiments, the action transfer module 706, during the process of executing the invocation of the artificial intelligence model corresponding to the action transfer sub-step for the action redirection sub-step, so as to execute the corresponding action transfer sub-step through the artificial intelligence model, is specifically used for: The reference motion video is parsed to obtain the reference skeleton topology information of the reference character, as well as the reference skeleton position information corresponding to each reference motion video frame. The reference skeleton topology information includes: the skeleton name of the reference character and the skeleton hierarchy relationship; the reference skeleton position information includes the skeleton position of the reference character. Input the reference skeleton topology information and the target skeleton topology information of the target character model into the skeleton mapping model, perform semantic alignment operation through the skeleton mapping model, and output the skeleton mapping relationship between the reference character and the target character model. The reference bone position information, bone mapping relationship and bone position of the target character model corresponding to each reference action video frame are input into the motion retargeting model. The motion retargeting model outputs the rotation angle sequence of each bone in the target character model. Based on the rotation angle sequence and the skinning weight values ​​of each vertex in the target character model, the positions of each bone and vertex in the target character model are dynamically adjusted to transfer the reference actions performed by the reference character to the target character model.

[0087] Optionally, in some embodiments, the video acquisition module 704 is specifically used for: Display action icons corresponding to different preset actions; in response to the selection operation of the target action icon, obtain the action video corresponding to the target action icon as a reference action video.

[0088] Optionally, in some embodiments, the video acquisition module 704 is specifically used for: Displays video upload guidance information; the video upload guidance information is used to guide the execution of the reference action video upload operation; receives the uploaded reference action video.

[0089] The motion transfer device of this embodiment is used to implement the corresponding motion transfer methods in the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here. Furthermore, the functional implementation of each module in the motion transfer device of this embodiment can be referred to the description of the corresponding part in the foregoing method embodiments, which will also not be repeated here.

[0090] Reference Figure 8 This document illustrates a schematic diagram of an electronic device according to an embodiment of this application. The specific embodiments of this application do not limit the specific implementation of the electronic device.

[0091] like Figure 8As shown, the electronic device may include: a processor 802, a communications interface 804, a memory 806, and a communications bus 808.

[0092] in: The processor 802, communication interface 804, and memory 806 communicate with each other through the communication bus 808.

[0093] Communication interface 804 is used for communication with other electronic devices.

[0094] The processor 802 is used to execute program 810, specifically to perform the relevant steps in the above-described action migration method embodiment.

[0095] Indicatively, program 810 may include program code that includes computer operation instructions.

[0096] The processor 802 may be a CPU, an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The smart device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0097] Memory 806 is used to store program 810. Memory 806 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0098] Program 810 may include multiple computer instructions. Specifically, program 810 can cause processor 802 to perform the operation corresponding to any of the methods described in the foregoing multiple method embodiments through multiple computer instructions.

[0099] The specific implementation of each step in program 810 can be found in the corresponding steps and units described in the above method embodiments, and has corresponding beneficial effects, which will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments, and will not be repeated here.

[0100] This application also provides a computer storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in any of the foregoing method embodiments. The computer storage medium includes, but is not limited to, compact disc read-only memory (CD-ROM), random access memory (RAM), floppy disk, hard disk, or magneto-optical disk.

[0101] This application also provides a computer program product, including computer instructions that instruct a computing device to perform an operation corresponding to any of the methods in the above-described multiple method embodiments.

[0102] Furthermore, it should be noted that the user-related information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to sample data used for training the model, data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0103] It should be noted that, depending on the implementation needs, the various components / steps described in the embodiments of this application can be broken down into more components / steps, or two or more components / steps or parts of the operation of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of this application.

[0104] The methods described in the embodiments of this application can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code downloaded over a network that is originally stored in a remote recording medium or a non-transitory machine-readable medium and will be stored in a local recording medium. Thus, the methods described herein can be stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an Application Specific Integrated Circuit (ASIC) or a Field Programmable Gate Array (FPGA)). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., Random Access Memory (RAM), Read-Only Memory (ROM), Flash Memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0105] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this application.

[0106] The above embodiments are only used to illustrate the embodiments of this application, and are not intended to limit the embodiments of this application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of this application. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of this application, and the patent protection scope of the embodiments of this application should be defined by the claims.

Claims

1. An action transfer method, comprising: Display the target character model; Get reference action videos; The reference action video includes a reference character performing the reference action; In response to a preset trigger operation, the AI ​​model corresponding to each action transfer sub-step is called sequentially according to the execution order of multiple action transfer sub-steps set in advance, so that each AI model executes the corresponding action transfer sub-step and transfers the reference action performed by the reference role to the target role model.

2. The method according to claim 1, wherein, After displaying the target character model, the method further includes: Display action migration process information; the action migration process information includes: the name information of each action migration sub-step, and the execution order information between each action migration sub-step.

3. The method according to claim 1 or 2, wherein, After displaying the target character model, the method further includes: Displays introductory information for the currently pending sub-step; the currently pending sub-step is the action transition sub-step with the earliest execution order among the unexecuted action transition sub-steps; The response to the preset trigger operation, in accordance with the pre-defined execution order of multiple action transition sub-steps, sequentially calls the artificial intelligence model corresponding to each action transition sub-step, including: In response to a trigger operation on the currently pending sub-step, the target artificial intelligence model corresponding to the currently pending sub-step is invoked, and the currently pending sub-step is executed through the target artificial intelligence model to obtain the execution result; According to the pre-defined execution order of multiple action migration sub-steps, the current sub-step to be executed is redefined, and the step of displaying the introductory information for the current sub-step to be executed is returned, until the reference action performed by the reference character is migrated to the target character model.

4. The method according to claim 1 or 2, wherein, The multiple action transfer sub-steps include, in sequence: a skeleton binding sub-step, a skinning weight optimization sub-step, and an action redirection sub-step.

5. The method according to claim 4, wherein, For the aforementioned skeleton binding sub-step, the process of invoking the AI ​​model corresponding to the action transfer sub-step, and executing the corresponding action transfer sub-step through the AI ​​model, includes: Obtain the semantic segmentation results for the target role model; The model data corresponding to the target character model and the semantic segmentation result are input into the skeleton recognition model, and the skeleton recognition model outputs the skeleton recognition result; the skeleton recognition result includes: joint position and joint name; Based on the joint positions and joint names, determine the bone positions and bone names of the target character model; Based on the bone positions and bone names of the target character model, the bone hierarchy of the target character model is determined; Based on the bone position, bone name, and bone hierarchy of the target character model, the skinning weights of each vertex in the target character model are initialized.

6. The method according to claim 4, wherein, For the skin weight optimization sub-step, the AI ​​model corresponding to the action transfer sub-step is invoked to execute the corresponding action transfer sub-step through the AI ​​model, including: Display the icons of the model parts corresponding to each model part in the target character model; In response to the selection operation of the target model part icon, obtain the target part data corresponding to the target model part and the bone position and bone name of each target bone contained in the target model part; The target part data, the bone position and bone name of each target bone are input into the skin weight optimization model, and the optimized skin weight of each vertex in the target model part is predicted by the skin weight optimization model.

7. The method according to claim 4, wherein, For the action redirection sub-step, the process of invoking the AI ​​model corresponding to the action transfer sub-step to execute the corresponding action transfer sub-step through the AI ​​model includes: The reference motion video is parsed to obtain the reference skeleton topology information of the reference character, and the reference skeleton position information corresponding to each reference motion video frame; the reference skeleton topology information includes: the skeleton name of the reference character and the skeleton hierarchy relationship; the reference skeleton position information includes the skeleton position of the reference character. The reference skeleton topology information and the target skeleton topology information of the target character model are input into the skeleton mapping model. The semantic alignment operation is performed through the skeleton mapping model, and the skeleton mapping relationship between the reference character and the target character model is output. The reference bone position information corresponding to each reference action video frame, the bone mapping relationship, and the bone position of the target character model are respectively input into the motion retargeting model, and the motion retargeting model outputs the rotation angle sequence of each bone in the target character model. Based on the rotation angle sequence and the skinning weight values ​​of each vertex in the target character model, the positions of each bone and vertex in the target character model are dynamically adjusted to transfer the reference action performed by the reference character to the target character model.

8. The method according to any one of claims 1-3, wherein, The acquisition of the reference action video includes: Display action icons corresponding to different preset actions; In response to a selection operation for a target action icon, the action video corresponding to the target action icon is obtained as a reference action video.

9. The method according to any one of claims 1-3, wherein, The acquisition of the reference action video includes: Display video upload guidance information; the video upload guidance information is used to guide the execution of the reference action video upload operation; Receive uploaded reference action videos.

10. A motion transfer device, comprising: The display module is used to display the target character model; The video acquisition module is used to acquire reference action videos; The reference action video includes a reference character performing the reference action; The action transfer module is used to respond to a preset trigger operation and sequentially call the artificial intelligence model corresponding to each action transfer sub-step according to the execution order of multiple preset action transfer sub-steps, so that each artificial intelligence model executes the corresponding action transfer sub-step and transfers the reference action performed by the reference character to the target character model.

11. An electronic device, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation corresponding to the method as described in any one of claims 1-9.

12. A computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any one of claims 1-9.

13. A computer program product comprising computer instructions that instruct a computing device to perform an operation corresponding to any one of the methods described in claims 1-9.

Images