Multi-character animation generation method for autistic children

Abstract

In order to solve the problems of role identity drift, action logic breakage, physical rule violation and the like in the existing generation technology, the invention discloses a multi-role animation generation method for a real scene of autistic children. A pre-training text-video diffusion model is embedded in a plug-and-play mode, and an attention mechanism of the pre-training text-video diffusion model is optimized on the premise of not retraining; in the generation process, the physical guidance module effectively guarantees the time sequence consistency and behavior rationality of role appearance, object movement and scene conversion, so that a stable animation conforming to a real world causal relationship is output; a predictable and reliable visual intervention material is constructed for autism children, and the ability of the autism children to understand social events and emotion expressions is significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of animation generation, in particular to a multi-role animation generation method for autistic children. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, generative models have made remarkable achievements in image and video generation, providing strong technical support for various application scenarios such as education, rehabilitation, and entertainment. In the field of special education, especially in the auxiliary teaching and emotional intervention for children with autism spectrum disorder, generative video content is considered an effective intervention tool with high visualization, strong interactivity, and realistic scenario simulation. Studies have shown that autistic children have certain difficulties in cognitive and understanding of social role interaction, and watching multi-role, contextually coherent, and clearly expressed visual media can help autistic children familiarize themselves with this aspect of ability.

[0003] Existing video generation methods have fundamental flaws when applied to autistic child intervention, as their generation logic is severely disconnected from real-world physical rules. The main performance is that in the generated content, the multi-role identity and spatial relationship are unstable, which destroys the consistency requirement of fixed teaching aids and social distance in intervention; the narrative logic jumps and the action breaks, which violates the causal chain followed by daily behaviors such as hand washing; the lack of physical constraints leads to object motion violating basic laws such as gravity, making it impossible to build a credible world model; at the same time, complex and subtle emotional expression also does not conform to the cognitive path of autistic children for exaggerated, single, and explicitly causal expressions. These problems collectively result in physically untrustworthy, logically incoherent, and cognitively mismatched generated videos, making them difficult to use for effective intervention.

[0004] On the other hand, in recent years, physical information neural networks (PINNs) have provided a new paradigm for modeling complex systems by incorporating physical priors such as control equations and initial boundary conditions, enabling supervised control of state evolution processes, as shown in the content of the patent with application publication number CN120563736A. However, traditional physical information neural networks are applied to numerical simulation or continuous field modeling, and in high-dimensional potential space image sequence generation tasks, they have high training costs and poor adaptability, and have not been effectively integrated into mainstream diffusion generation frameworks. Therefore, a lightweight framework that can be embedded in existing generation models is needed. SUMMARY

[0005] The purpose of the present application is to solve the problems of the prior art and provide a multi-role animation generation method for autistic children.

[0006] To solve the above problems, the present application adopts the following technical solutions: The application discloses a multi-role animation generation method for autistic children, and comprises the following steps: Step S1: constructing intervention targets and input conditions, converting input teaching intention instructions into animation generation instructions through a multi-modal information fusion technology; Step S2: loading a pre-trained text-to-video generation model as a basic generator , and freezing model parameters; Step S3: integrating a physical information attention module, embedding physical dynamics constraints into the attention mechanism of the pre-trained text-to-video diffusion model according to visual stability requirements; Step S4: generating animations by performing physical regularization constraints, integrating the physical information attention module in S3 into the actual sampling process of the basic generator in S2 , and then performing an improved sampling algorithm based on physical constraints to generate final video data.

[0007] Further, the process of constructing intervention targets and input conditions in the step S1 comprises three steps, namely, a structured text prompt word engineering, a preparation and processing of a starting condition image and a joint encoding of condition inputs. The structured text prompt word engineering means converting preset natural language instructions about animation content into structured prompt words. The preparation and processing of the starting condition image means fixing the initial state of the generated content, the appearance of a role and the layout of a scene, and enhancing the consistency of the generation starting point. The joint encoding of the condition inputs means jointly encoding the text and image conditions into a unified condition vector as the input of the generation model.

[0008] Further, the process of the preparation and processing of the starting condition image firstly provides a starting condition image , which is generated by a pre-trained text-to-image model according to the local prompt words obtained in the structured text prompt word engineering ; and then, a role mask in the image is extracted.

[0009] Further, the process of the joint encoding of the condition inputs comprises the following steps: Step S11: text encoding, using a pre-trained text encoder to map the global prompt and the local prompt sequence obtained in the structured text prompt word engineering into feature vector sequences respectively: , wherein, ​denotes a pre-trained CLIP model-based text encoder; Step S12: image encoding, using a pre-trained visual encoder to encode the starting image and the mask of the image to obtain visual features and spatial features ; Step S13: integrating the multi-modal features through a fusion network to obtain a conditional vector as follows: .

[0010] Further, the CogVideoX model pre-trained in step S2 is selected as the base generator ; the model works through a reverse process of gradually denoising a noise to generate a video ; each step of the reverse denoising process predicts a noise based on a conditional vector , and the mean of the noise is parameterized by the following formula: , wherein, denotes the variance of the noise added to the data at time step t; is , denoting the proportion of the signal at time step t that is preserved from the signal at the previous time step without being destroyed by the noise; is the product of all from the first step to the t-th step; denotes a denoising network comprising spatial, temporal and cross-attention layers.

[0011] Further, the process of integrating the physical information attention module in step S3 comprises the following steps: Step S31: physical dynamics modeling; Step S32: incorporating the physical constraints obtained in step S31 into the calculation of the attention weights of the temporal attention layer of the text-to-video diffusion model in step S2; Step S33: performing adaptive adjustment of physical parameters.

[0012] Further, the process of physical dynamics modeling in step S31 comprises: first, representing the feature of the video sequence in the latent space as , wherein is the latent feature of the t-th frame; and establishing a physical prior model based on the heat conduction equation: , where, is the thermal diffusivity, is the Laplacian operator, is the latent feature; using the discretized heat conduction equation, from the latent feature of the current frame predict the feature of the next frame : , where, is the discretized Laplacian operator matrix; is the time step; is the physically predicted next frame feature; is the latent feature of the t-th frame.

[0013] Further, in the step S32, the modified physical guidance attention is represented as: , where, is the concatenation function; are the query projection matrix, the key projection matrix, and the value projection matrix, respectively; is the dimension of the Key, and the scaling factor is used to prevent the numerical value of the dot product from being too large; In the diffusion back-sampling process of the diffusion model from text to video described in step S2, the generation direction is guided by the physical gradient; the score function in the modified diffusion model is represented as: , where, is the guidance intensity coefficient; represents the physical loss; is the gradient operation on the latent feature of the current frame ; The physical constraint is applied on different resolution levels of the U-Net of the model to perform multi-scale physical constraint: , where, is the level weight, is the level-related physical loss.

[0014] Further, in the step S4, the physical regularization generation process is performed: After introducing the physical constraint in S3, the physical guidance sampling algorithm is performed; it is represented as: , where, is the score function corrected by physical gradient, representing the noise prediction of the model; physical guidance coefficient decays over time, represented as: , where the hyperparameter controls the decay rate of the physical constraint; Secondly, multi-scale feature constraint is performed, in each step of sampling, from different resolution levels of the U-Net of the model in S2 feature maps are extracted ; based on the feature maps , the multi-scale calculation is performed on the physical loss function defined in S3, represented as: , where the calculation of the physical loss of each layer is consistent with the physical dynamics model defined in S3, and the hierarchical weight coefficient is assigned according to the importance of the U-Net layer; In addition, an adaptive guidance mechanism needs to be established, and the physical guidance strength is adaptively adjusted according to the current sampling state: , where is a small constant set; is the physical guidance coefficient at the initial moment; Finally, timing consistency monitoring needs to be performed, and the timing consistency index of the subsequent step S5 is monitored in real time during the sampling process When is lower than the threshold , the physical constraint strength is temporarily enhanced, represented as: , where is an enhancement coefficient.

[0015] Further, it further includes step S5: evaluation and output of the generated animation, including evaluation of three dimensions of visual quality, physical consistency and cognitive suitability of the generated animation; visual quality includes frame-level quality and video-level quality; physical consistency includes time stability and time drift; cognitive suitability includes intervention intention alignment, semantic consistency and role identity consistency.

[0016] The beneficial effects of the present application are: By utilizing a pre-trained basic generator without introducing additional training burden, and through parameter improvements, injecting character semantic attributes, strengthening temporal dependencies, and introducing physical common sense constraints, the generated animated videos are ensured to be highly consistent in terms of identity, timing, and actions. This builds a stable and predictable visual cognitive framework for children with autism, thereby significantly improving their ability to understand events, interactions, and emotions. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-role animation generation method for real-life scenarios of children with autism according to the present invention; Figure 2 This is a diagram of the animation generation architecture of the present invention. Detailed Implementation

[0018] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0019] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the figures only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] Example 1:

[0021] like Figure 1 and Figure 2 As shown, a method for generating multi-character animations for children with autism is based on integrating a training-free physical information attention module into a pre-trained large-scale video diffusion model, thereby imposing physical constraints during the generation process. The specific implementation methods for each step are as follows: Step S1 involves constructing the intervention goals and input conditions. This step is fundamental to generating animation. Its goal is to transform abstract educational intervention goals into structured conditional signals that the generative model can accurately understand and execute. These educational intervention goals are typically well-structured and semantically clear descriptions of teaching scenarios written by educational experts in the field, focusing on interactions between social roles. This step includes three levels: structured text cue engineering, preparation and processing of initial conditional images, and joint encoding of conditional inputs.

[0022] The structured text prompting project specifically refers to guiding the model to generate content that aligns with the cognitive characteristics of children with autism. This requires converting pre-defined natural language instructions regarding the animation content into structured prompts. For example, the final generated video sequence is defined as follows: ,in The total number of frames; the global text prompts corresponding to this video sequence are denoted as... Describe the core narrative of the entire video, such as: "Xiaoming and Xiaohong are learning to take turns playing soccer on the playground."

[0023] To achieve more precise timing control, global text prompts will be added. Decomposed into a sequence of local cue words corresponding to video segments or keyframes, represented as ,in, This means that each image frame in the video corresponds to at most one local cue word. Each local cue word describes a specific sub-event or state transition. For example, the global text cue "Xiaoming and Xiaohong are learning to take turns playing soccer on the playground" can be broken down into examples as shown in Table 1 below: Table 1. Example of Decomposition

[0024] Decomposing global text prompts locally provides the model with a clear temporal structure and causal logic clues, helping to generate logically coherent animation effects.

[0025] The preparation and processing of the initial condition image specifically refers to fixing the initial state of the generated content, such as the character appearance and scene layout, to enhance the consistency of the generation starting point. For example, providing an initial condition image. The image A pre-trained text-to-image model can be used to identify local cue words. generate.

[0026] Subsequently, the image Preprocessing is performed to extract stronger control signals; in this case, preprocessing includes character mask extraction. Character mask extraction requires automatically acquiring images using an instance segmentation model. The binary mask for each character is shown below:

[0027] in, Indicates the first The pixel regions where each role is located; SAM() represents the SAM instance segmentation algorithm. These role masks provide precise spatial guidance for maintaining role identity consistency in subsequent steps.

[0028] The joint encoding of the condition input is specifically represented as jointly encoding the text and image conditions into a unified condition vector As the input of the generation model, it specifically includes the following steps: First, text encoding is performed, and the global prompt and the local prompt sequence are respectively mapped into a sequence of feature vectors using a pre-trained text encoder of a text encoder: ,

[0029] wherein represents a pre-trained text encoder based on a CLIP model; Second, image encoding is performed, and the starting image and the mask of the image are respectively encoded using a pre-trained visual encoder to obtain visual features and spatial features ; Finally, feature fusion is completed, and the multi-modal features are integrated through a fusion network to obtain a condition vector , as shown below: ,

[0030] wherein an information-rich and structured condition vector is output, which provides an accurate guide blueprint for subsequent physical regularization animation generation, ensuring that the generated content is highly aligned with the input global text prompt from the beginning, guaranteeing specific intervention targets for autistic children, and providing a solid foundation for the model to maintain spatio-temporal consistency.

[0031] Step S2 loads a pre-trained text-to-video diffusion model as a frozen base generator In this example, the CogVideoX model is selected; this model works through a reverse process of gradually denoising a noise to generate a video , and the core of its denoising network is a U-Net architecture containing spatial, temporal, and cross-attention layers. After loading the pre-trained diffusion model, all its parameters are completely frozen, i.e. to preserve its prior knowledge and reduce the burden of additional training.

[0032] Each step of this reverse denoising process is based on the condition vector to predict the noise , and its mean is parameterized by the following formula: , where, is the variance of the noise added to the data at time step t; is defined as 1- , represents the proportion of the signal at time step t that is preserved from the previous time step, rather than being destroyed by noise; is the product of all from step 1 to step t; denotes the denoising network containing spatial, temporal and cross-attention layers. The conditional vector and the latent feature noise provide a unified input interface for subsequent introduction of physical constraints.

[0033] Step S3 integrates the physical information attention module, which realizes the improvement of the spatio-temporal consistency of the generated content by embedding the physical dynamics constraints into the attention mechanism of the pre-trained text-to-video diffusion model. The specific implementation is as follows: The physical dynamics modeling represents the feature of the video sequence in the latent space as where is the latent feature of the t-th frame. The physical prior model is established based on the heat conduction equation: , where, is the thermal diffusivity, is the Laplacian operator, is the latent feature.

[0034] The next frame feature is predicted from the current frame latent feature using the discretized heat conduction equation: , where, is the discrete Laplacian matrix; is the time step; here is the physically predicted next frame feature, which can be regarded as a "rough estimate" of the next frame feature based on the physical prior; is the latent feature of the t-th frame.

[0035] In the temporal attention layer of the text-to-video diffusion model described in S2, the above physical constraints are integrated into the attention weight calculation. The standard attention is calculated as: , where, denotes the target that needs to be focused on at present, corresponding to Query; Identity tag or feature representing input data, corresponding to Key; Actual content of input data, corresponding to Value; Dimension of Key, scaling factor Used to prevent the gradient problem caused by the large dot product value; In this example, when calculating the time attention, in order to make the model find the relationship between and more quickly and efficiently, the parameter is added to guide the relationship between the two. The modified physical guidance attention is: , wherein, is splicing, aiming to fuse the current frame and the physical prediction, is query projection, used for the current frame; , , Query projection matrix, key projection matrix, and value projection matrix, respectively.

[0036] In the diffusion reverse sampling process of the text-to-video diffusion model described in step S2, the generation direction is guided by the physical gradient; the score function in the modified diffusion model is represented as: , wherein, is the guidance intensity coefficient; represents the physical loss; represents the gradient operation on the latent feature of the current frame.

[0037] The physical constraint is applied on different resolution levels of the U-Net of the model to perform multi-scale physical constraint: , wherein, is the level weight, is the physical loss related to the level.

[0038] Finally, the physical parameter adaptive adjustment mechanism is performed, as shown in the following formula: , wherein, is the decay coefficient, is the denoising step number, is the physical parameter adaptive adjustment coefficient set at the initial moment.

[0039] In this step, the time attention layer of the text-to-video diffusion model integrated into step S2 through the forward hook is improved significantly with a parameter increment of less than 1% of the base model, which improves the time continuity of the generated video, and the adjusted model can be applied to autism intervention materials and other scenes with strict stability requirements.

[0040] In step S4, the physical regularization generation process is performed, including integrating the physical information attention module in S3 into the base generator in S2 The actual sampling process of the base generator in S2 is as follows: first, a physical constraint-based improved sampling algorithm is performed, starting from pure noise , and through times of iteration, the final video data is generated.

[0041] The physical regularization generation process is first performed by a physical guided sampling algorithm; wherein in the standard sampling step, the generation process of the t-th step is: , wherein represents the random noise of the probability distribution in the reverse sampling process.

[0042] After introducing the physical constraint in S3, the sampling process is modified as: , wherein is the score function modified by the physical gradient, and represents the noise prediction of the model. The physical guide coefficient decays over time and is represented as: , wherein the hyperparameter controls the decay rate of the physical constraint, and the animation is applied with strong constraint to stabilize the structure in the early stage of generation, and gradually weakened to retain details in the later stage; is the physical guide coefficient set at the initial time.

[0043] Secondly, multi-scale feature constraint is performed, in each step of sampling, feature maps are extracted from different resolution levels of the U-Net of the model in S2 ; based on the feature maps , the physical loss function defined in S3 is calculated in multiple scales, represented as: , wherein the calculation of the physical loss of each layer is consistent with the physical dynamics model defined in S3, and the layer weight coefficient is assigned according to the importance of the U-Net layer.

[0044] In addition, an adaptive guidance mechanism needs to be established to adjust the physical guidance strength According to the current sampling state, adjust adaptively: , where, is a small constant set to prevent division by zero; is the initial physical guidance coefficient. This mechanism is used to ensure the balance between physical constraints and semantic guidance.

[0045] Finally, timing consistency monitoring needs to be performed, and the timing consistency index of the subsequent step S5 is monitored in real time during the sampling process When is lower than the threshold , temporarily increase the physical constraint strength, denoted as: , where, is the enhancement coefficient. This mechanism can intervene immediately when flickering risk is detected, ensuring the stability of the generated animation.

[0046] Step S5 is the evaluation and output of the generated animation, which evaluates the animation sequence generated in step S4 in multiple dimensions to ensure that it meets the technical requirements of autism intervention. The evaluation system is shown in Table 2 below, which covers three dimensions of visual quality, physical consistency, and cognitive suitability, aiming to comprehensively verify the performance of the generated content in terms of technical advancement and intervention effectiveness.

[0047] Table 2 Comprehensive evaluation index system of generated animation

[0048] For the parameter description in Table 2, among them, the frame-level quality FID calculation formula is: is the mean of the real data feature vector. is the mean of the generated data feature vector. This part calculates the square of the Euclidean distance between the means of the two distributions. It measures the deviation of the center points of the two distributions. is the covariance matrix of the real data feature vector. is the covariance matrix of the generated data feature vector. Secondly, the time stability TSI calculation formula is: is the entire generated video sequence. is a pre-trained feature extraction function, the input is a frame of image, and the output is a feature vector. In addition, the intervention intention alignment degree calculation formula is: is a video classification model pre-trained on the autism intervention behavior dataset, used to identify whether a specific intervention element is included in the video; is the binary label vector parsed from the prompt word C according to S1, indicating which intervention elements the video should theoretically contain; is an indicator function; is the total number of intervention elements, which directly measures the ability of the system to convert abstract intervention instructions into specific and identifiable intervention visual content. Finally, the formula of the role identity consistency IDConsist, is the t-th frame image, is the reference frame image, is a feature extraction function, is a cosine similarity function.

[0049] Finally, a comprehensive score function is established to comprehensively evaluate the intervention effect, and the newly added intervention intention alignment IIA should be given a significant weight: , where the weight coefficient to is adjusted adaptively according to the evaluation feedback of autism experts; it should be noted that Among the five weights, a higher level is maintained, exceeding the average weight, to ensure that the evaluation results are highly consistent with the core intervention goals.

[0050] Through this multi-dimensional evaluation system, especially the "intervention intention alignment" indicator, the end-to-end efficiency of the invention from intervention instruction input to effective intervention content generation can be directly verified, ensuring that the generated animation content is above the required technical quality and meets the strict requirements of autism intervention.

[0051] The above description is only one specific example of the present application and does not constitute any limitation on the present application. Obviously, for those skilled in the art, after understanding the content and principles of the present application, various modifications and changes in form and details can be made without departing from the principles and structures of the present application, but these modifications and changes based on the idea of the present application are still within the protection scope of the claims of the present application.

Claims

1. A method for generating multi-character animations for children with autism, characterized in that, Includes the following steps: Step S1: Construct intervention goals and input conditions, and transform the input teaching intention instructions into animation generation instructions through multimodal information fusion technology; Step S2: Load the pre-trained text-to-video generation model as the base generator. And freeze the model parameters; Step S3: Integrate the physical information attention module and, based on the visual stability requirements, embed physical dynamic constraints into the attention mechanism of the pre-trained text-to-video diffusion model; Step S4: Execute physical regularization constraints to generate animation, integrating the physical information attention module in S3 into the base generator in S2. In the actual sampling process, an improved sampling algorithm based on physical constraints is then executed to generate the final video data.

2. The method for generating multi-character animations for children with autism according to claim 1, characterized in that, The process of constructing the intervention target and input conditions in step S1 includes three steps: structured text prompt word engineering, preparation and processing of the initial condition image, and joint encoding of the condition input. The structured text prompting technology transforms pre-defined natural language instructions about animation content into structured prompts. The preparation and processing of the initial condition image represents the initial state of the generated content, including the character appearance and scene layout, thus enhancing the consistency of the generation starting point. The joint encoding of conditional inputs represents the joint encoding of text and image conditions into a unified conditional vector. , as input to the generative model.

3. The method for generating multi-character animations for children with autism according to claim 2, characterized in that, The process of preparing and processing the initial condition image begins by providing an initial condition image. The image Local cues obtained from structured text cue engineering are obtained through a pre-trained text-to-image model. Generate; then, extract the image. The role mask in the middle.

4. The method for generating multi-character animations for children with autism according to claim 3, characterized in that, The joint encoding process of the conditional input includes the following steps: Step S11: Perform text encoding, using a pre-trained text encoder to encode the global prompts obtained from the structured text prompt word engineering. and local cue sequences Each is mapped to a sequence of feature vectors: ， in, This represents a pre-trained CLIP-based text encoder. Step S12: Perform image encoding, using a pre-trained visual encoder to encode the starting image. and image mask Encode to obtain visual features and spatial features ; Step S13: Through a converged network The multimodal features are integrated to obtain the conditional vector. As shown below: 。 5. A method for generating multi-character animations for children with autism according to claim 4, characterized in that, In step S2, a pre-trained CogVideoX model is selected as the base generator. The model uses a noise-driven approach. Gradual denoising to generate video The reverse process is used; each step of the reverse denoising process is based on condition vectors. To predict noise Its mean Parameterized by the following formula: ， in, This represents the variance of the noise added to the data at time step t; for This indicates that at time step t, the signal from the previous time step... The proportion that is preserved rather than destroyed by noise; It is all from step 1 to step t. The product of consecutive products; This represents a denoising network that includes spatial, temporal, and cross-attention layers.

6. The method for generating multi-character animations for children with autism according to claim 1, characterized in that, The process of integrating the physical information attention module in step S3 includes the following steps: Step S31: Physical dynamics modeling; Step S32: Integrate the physical constraints obtained in step S31 into the attention weight calculation of the temporal attention layer of the text-to-video diffusion model described in step S2; Step S33: Perform adaptive adjustment of physical parameters.

7. A method for generating multi-character animations for children with autism according to claim 6, characterized in that, The physical dynamics modeling process in step S31 includes: first, representing the features of the video sequence in the latent space as... ,in The potential features of frame t; a physical prior model is established based on the heat conduction equation: ， in, Where is the thermal diffusivity, For the Laplace operator, As a potential feature; Using the discretized heat conduction equation, from the latent features of the current frame Predicting features in the next frame : ， in, It is a discrete Laplace operator matrix; For time step; Features for the next frame predicted by physics; Let be the potential features of the t-th frame.

8. A method for generating multi-character animations for children with autism according to claim 7, characterized in that, In step S32, the modified physical guidance attention is first expressed as follows: ， in, This is a concatenation function; These are the query projection matrix, key projection matrix, and value projection matrix, respectively. It is the dimension of the key, the scaling factor. Used to prevent the dot product value from being too large; In the backsampling process of the text-to-video diffusion model described in step S2, the generation direction is guided by physical gradients; the score function in the modified diffusion model is expressed as: ， in, The guiding strength coefficient; Indicates physical loss; For the latent features of the current frame Gradient calculation; At different resolution levels of the U-Net in the model Apply physical constraints to perform multi-scale physical constraints: ， in, For hierarchical weights, This refers to the physical loss related to the hierarchy.

9. A method for generating multi-character animations for children with autism according to claim 8, characterized in that, In step S4, a physical regularization generation process is performed: First, after introducing the physical constraints in S3, a physical guided sampling algorithm is performed; represented as: ， in, It is the score function corrected for physical gradients, representing the model's noise prediction. This represents random noise in the probability distribution during backsampling; physical guidance coefficient. Decay over time, expressed as: ， Among them, hyperparameters Control the decay rate of physical constraints; Secondly, multi-scale feature constraints are implemented. In each sampling step, the U-Net of the model described in S2 is sampled at different resolution levels. Extract feature maps Based on feature maps The physical loss function defined in S3 is calculated on a multi-scale basis and expressed as follows: ， Among them, the physical loss of each layer The calculation is consistent with the physical dynamics model defined in S3, with hierarchical weight coefficients. Assigned according to the hierarchical importance of U-Net; In addition, an adaptive guidance mechanism and physical guidance strength need to be established. Adaptively adjust based on the current sampling status: ， in, A small constant that is set; The physical guidance coefficient at the initial moment; Finally, timing consistency monitoring is required, which involves real-time monitoring of the timing consistency metrics in subsequent step S5 during the sampling process. ,when Below the threshold When the physical constraint strength is temporarily increased, it is expressed as: ， in, This is the enhancement coefficient.

10. A method for generating multi-character animations for children with autism according to claim 1, characterized in that, It also includes step S5: evaluation and output of the generated animation, including evaluation of the generated animation in three dimensions: visual quality, physical consistency and cognitive suitability; visual quality includes frame-level quality and video-level quality; physical consistency includes temporal stability and temporal drift; cognitive suitability includes the alignment of intervention intention, semantic consistency and role identity consistency.

Images