Identity-enhanced figure image animation generation method based on diffusion model

Abstract

An identity-enhanced figure image animation generation method based on a diffusion model specifically relates to the technical field of figure image animation generation, and comprises an image acquisition module, an image expansion extraction module and an image processing module, camera equipment of the image acquisition module acquires an original figure image and M original features of the original figure image, the diffusion model training unit adds diffusion operation to an original figure image step by step to obtain a first feature variation, the image processing module obtains a first animation image, the feature extraction and evaluation unit evaluates a local feature error of the first animation image to obtain a correction coefficient matrix, and the image processing module obtains a second animation image. The image processing module obtains a third animation image and transmits the third animation image to the diffusion model training unit, identity-enhanced figure image animation generation based on the diffusion model can be realized, and the generated animation sequence is continuously and circularly optimized, so that the continuity of adjacent frames in feature and quality is further ensured.

Description

Technical Field

[0001] This invention relates to the field of character image animation generation technology, specifically a method for generating character image animation with identity enhancement based on a diffusion model. Background Technology

[0002] In recent years, diffusion models have made significant progress in the fields of image and video generation. In image generation tasks, diffusion models learn to reverse the process of gradually adding noise to images, enabling the generation of high-quality samples from complex distributions. In video generation, many studies have drawn on the structure of text-to-image models, transforming pre-trained text-to-image models into video generators by enhancing inter-frame attention modeling. In the field of human image animation, diffusion models are widely used due to their superior generation quality and stable controllability.

[0003] Currently, some video generation methods based on diffusion models do not focus on identity preservation by default, which can easily lead to identity drift and background inconsistencies. In particular, existing methods have problems in handling texture inconsistencies and temporal instability, resulting in a lack of research on generalization ability in character animation. Moreover, existing technologies may not be able to accurately correct the errors between local features and expected features in the initial animation image. Finally, existing technologies are not accurate enough in evaluating the overall image quality, making it impossible to reasonably adjust the intensity of the subsequent forward diffusion.

[0004] Therefore, those skilled in the art have provided a method for generating identity-enhanced character image animations based on a diffusion model to address the problems mentioned in the background art. Summary of the Invention

[0005] The technical problem solved by this invention is to provide a method for generating character image animations with enhanced identity based on a diffusion model, so as to achieve enhanced identity features, accurate correction of local feature errors and overall quality optimization, thereby achieving animation continuity.

[0006] To address the above problems, the present invention provides the following technical solution:

[0007] A method for generating identity-enhanced character images and animations based on a diffusion model includes an image acquisition module, an image expansion and extraction module, and an image processing module. The image expansion and extraction module includes a diffusion model training unit and a feature extraction and evaluation unit.

[0008] The specific generation method and steps are as follows:

[0009] Step 1: Use the camera device of the image acquisition module to acquire the original image of the person and M original features of the original image of the person, and transmit them to the image processing module;

[0010] Step 2: Using the diffusion model training unit, Gaussian noise is gradually added to the original human image on the computer to obtain the first feature change amount and transmit it to the image processing module.

[0011] Step 3: Obtain from the image processing module the first feature average reference, the correction coefficient matrix reference, the second feature average reference, the weight coefficient group, and the diffusion coefficient group that sets the diffusion value rules.

[0012] Step 4: The image processing module obtains the first animation image initially generated after forward diffusion based on the original character image, the first feature change amount, and the diffusion coefficient group, and transmits it to the feature extraction and evaluation unit.

[0013] Step 5: The feature extraction and evaluation unit evaluates the local feature error of the first animation image, obtains the correction coefficient matrix, and transmits it to the image processing module.

[0014] Step 6: The image processing module obtains the second animation image after back-diffusion adjustment based on the correction coefficient matrix, the first animation image, and the diffusion coefficient group, and transmits them to the image expansion extraction module respectively.

[0015] Step 7: Using the diffusion model training unit and the feature extraction and evaluation unit, obtain the second feature change and the optimization coefficient matrix respectively, and transmit them to the image processing module;

[0016] Step 8: The image processing module obtains a third animation image based on the second animation image, the second feature change of the diffusion coefficient group, and the optimization coefficient matrix, and transmits it to the diffusion model training unit.

[0017] Repeat steps 2-8 to complete the overall animation sequence.

[0018] Furthermore, the diffusion coefficient set includes the primary forward diffusion coefficient, the reverse diffusion coefficient, and the secondary forward diffusion coefficient.

[0019] Further: The initial positive expansion coefficient is multiplied by the first feature change to obtain the initial positive adjustment amount of the original human image during the positive diffusion stage;

[0020] The first animated image is obtained by adding the initial positive adjustment amount to the original character image.

[0021] Furthermore: the weight coefficient group includes a first feature weight reorganization and a second feature weight reorganization;

[0022] The image processing module uses the arithmetic mean method to obtain the initial feature average value based on the M original features and the first feature weighting.

[0023] Based on the comparison between the initial feature average and the first feature average benchmark, the value of the initial positive expansion coefficient is obtained. The specific diffusion value rules for the initial positive expansion coefficient are as follows:

[0024] If the initial feature average value is greater than or equal to the first feature average benchmark, then the initial positive expansion coefficient is 0.7.

[0025] If the initial feature average value is less than the first feature average benchmark, then the initial positive expansion coefficient is 0.3.

[0026] Further: Subtract the original character image from the first animated image to obtain a forward diffusion variable that reflects the changes in the image during the forward diffusion process;

[0027] The forward diffusion variable is multiplied by the reverse expansion coefficient to obtain the adjustment base quantity;

[0028] The adjustment base amount is multiplied by the correction coefficient matrix to obtain the back diffusion adjustment amount;

[0029] The second animation image is obtained by subtracting the back diffusion adjustment amount from the first animation image;

[0030] The specific steps for obtaining the correction coefficient matrix are as follows:

[0031] Step 1: Extract single-region features from the local area of ​​the first animation image, including features of the eyes and mouth regions of the face;

[0032] Step 2: Compare the extracted single-region features with the expected features to obtain the feature error between the two;

[0033] Step 3: Generate the correction coefficient matrix of the local region based on the comparison results of all the feature errors in the local region.

[0034] Furthermore: the specific diffusion value rules for the reverse expansion coefficient, based on the correction coefficient matrix and the reference of the correction coefficient matrix, are as follows:

[0035] If the correction coefficient matrix is ​​greater than the correction coefficient matrix base, then the reverse expansion coefficient is 0.6;

[0036] If the correction coefficient matrix is ​​smaller than the correction coefficient matrix base, then the reverse expansion coefficient is 0.4.

[0037] Further: The diffusion model training unit gradually adds Gaussian noise to the original human image on a computer to obtain the second feature change amount;

[0038] The second positive expansion coefficient is multiplied by the second characteristic change to obtain the second positive adjustment amount after the second positive adjustment;

[0039] The secondary positive adjustment amount is multiplied by the optimization coefficient matrix to obtain the overall adjustment amount after considering the overall image quality.

[0040] The second animated image is combined with the aforementioned comprehensive adjustment amount to obtain the third animated image.

[0041] Furthermore: the image acquisition module works in conjunction with the image expansion and extraction module, and uses the arithmetic mean method to obtain M in the second animation image. ’ A quadratic positive post-feature;

[0042] The image processing module uses the arithmetic mean method to obtain the result based on M. ’ The average value of the quadratic features obtained by weighting the second feature and the quadratic positive backward features.

[0043] Further: Based on the second feature average benchmark and the quadratic feature average, the specific diffusion value rules for the quadratic positive expansion coefficient are as follows:

[0044] If the average value of the quadratic features is greater than or equal to the second feature average benchmark, then the value of the quadratic positive expansion coefficient is 0.2.

[0045] If the average value of the quadratic features is less than the average benchmark of the second features, then the value of the quadratic positive expansion coefficient is 0.3.

[0046] The effects of the above solution are as follows:

[0047] 1. This invention introduces an initial forward expansion coefficient into the diffusion coefficient group, which can be adjusted according to the actual situation of the image quality and the degree of obviousness of identity features presented by comparing the average value of the initial features with the average value of the first feature benchmark. Based on the diffusion value rules, the initial forward expansion coefficient is adjustable, and the forward diffusion process is more controllable, avoiding the image distortion problem caused by excessive diffusion. It ensures that the original image features can be changed while the identity features of the person can be maintained during the forward diffusion stage.

[0048] 2. The present invention uses the correction coefficient matrix obtained by the feature extraction and evaluation unit to compare the single-region features in a local area of ​​the original human image with the expected features, calculates the error using the mean square error method, and thus accurately locates the local feature error. Then, combined with the back diffusion coefficient, the preliminary animation image is adjusted, which effectively solves the problem of insufficient local feature error correction in the prior art and makes the image closer to the identity features of the original image.

[0049] The back diffusion coefficient is adjusted based on the local feature error and the actual situation of the overall image stability presented by the comparison between the correction coefficient matrix and the benchmark correction coefficient matrix. This avoids over-adjustment that could lead to overall image instability and improves the stability of the image during the back diffusion adjustment stage.

[0050] 3. The optimization coefficient matrix of the present invention comprehensively considers multiple features of image clarity, contrast, and color saturation, and performs an overall quality assessment of the second animation image. Then, combined with the quadratic positive expansion coefficient, the second animation image is further optimized, which solves the problem of inaccurate overall quality assessment in the prior art, thereby further improving the quality of the generated animation.

[0051] Among them, the second positive expansion coefficient is adjusted according to the actual situation of the overall image quality and animation continuity requirements presented by the comparison of the second feature average benchmark and the second feature average value, so as to better adapt to the dynamic changes of animation, solve the problem of insufficient texture and time stability in the existing technology, and enhance the continuity and stability between animation frames. Attached Figure Description

[0052] Figure 1 The overall method for generating character image animations to enhance this identity involves the following steps:

[0053] Figure 2 This is a schematic diagram illustrating the diffusion adjustment of an image by the diffusion coefficient group in this invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0055] Example 1, please refer to Figure 1-2 An identity enhancement character image animation generation method based on diffusion model includes an image acquisition module, an image expansion and extraction module, and an image processing module. The image expansion and extraction module includes a diffusion model training unit and a feature extraction and evaluation unit.

[0056] The specific generation method and steps are as follows:

[0057] Step 1: Use the camera device of the image acquisition module to acquire the original image of the person and M original features of the original image of the person, and transmit them to the image processing module;

[0058] Step 2: Using the diffusion model training unit, Gaussian noise is gradually added to the original human image on the computer to obtain the first feature change and transmit it to the image processing module.

[0059] Step 3: Obtain from the image processing module the first feature average benchmark, the correction coefficient matrix benchmark, the second feature average benchmark, the weight coefficient group, and the diffusion coefficient group that sets the diffusion value rules.

[0060] The diffusion coefficient group includes the primary forward expansion coefficient, the reverse expansion coefficient, and the secondary forward expansion coefficient; the weight coefficient group includes the first feature weight recombination and the second feature weight recombination.

[0061] Step 4: The image processing module obtains the first animation image initially generated after forward diffusion based on the original human image, the first feature change amount and the diffusion coefficient group, and transmits it to the feature extraction and evaluation unit.

[0062] Step 5: The feature extraction and evaluation unit evaluates the local feature error of the first animation image, obtains the correction coefficient matrix, and transmits it to the image processing module.

[0063] Step 6: The image processing module obtains the second animation image after back-diffusion adjustment based on the correction coefficient matrix, the first animation image, and the diffusion coefficient group, and transmits them to the image expansion and extraction module respectively.

[0064] Step 7: Using the diffusion model training unit and the feature extraction and evaluation unit, obtain the second feature change and the optimization coefficient matrix respectively, and transmit them to the image processing module;

[0065] Step 8: The image processing module obtains the third animation image based on the second animation image, the second feature change of the diffusion coefficient group, and the optimization coefficient matrix, and transmits it to the diffusion model training unit.

[0066] Repeat steps 2-8 to complete the entire animation sequence;

[0067] In this embodiment, the optimization of the original character image to the third animation image during the identity enhancement process based on the diffusion model can generate a series of continuous, high-quality character image animation frames with well-preserved identity features.

[0068] After processing one frame and obtaining a better third animation image, the generation of the next frame does not require reusing the original character image. Instead, the optimized third animation image from the previous frame is used as the new original character image and substituted into the diffusion model training unit to restart the forward diffusion process. This cycle repeats, ensuring that each iteration is based on the optimized result of the previous frame. This results in better continuity in features and quality between adjacent frames. Since the optimized third animation image from the previous frame already has a certain degree of good identity feature preservation and animation effect, using the third animation image as a new starting point allows the next frame to be further optimized and adjusted on that basis, thereby gradually accumulating optimization effects throughout the entire animation sequence.

[0069] Please see Figure 1-2 The initial positive expansion coefficient is multiplied by the change in the first feature to obtain the initial positive adjustment amount of the original human image during the positive diffusion stage.

[0070] The original character image is adjusted with an initial positive adjustment to obtain the first animation image;

[0071] The image processing module uses the arithmetic mean method to obtain the initial feature average value based on the M original features and the first feature weight reorganization.

[0072] The formula for calculating the first animation image in this embodiment is as follows:

[0073] T A =T RAW +t1×TB RAW ;

[0074] in:

[0075] T A The first animated image;

[0076] T RAW Original human image;

[0077] t1 is the initial positive expansion coefficient;

[0078] TB RAW This is the change in the first characteristic;

[0079] t1×TB RAW The calculation result is the initial positive adjustment amount.

[0080] The formula for calculating the initial characteristic mean is as follows:

[0081]

[0082] in:

[0083] T avgThe initial characteristic average;

[0084] M represents the total amount of original features;

[0085] T j Let a be the j-th feature among M original features. j It represents the j-th weight in the first feature weight reorganization.

[0086] This algorithm unit is in the initial positive diffusion stage and utilizes t1×TB RAW Calculate the initial positive adjustment amount, and then adjust its relationship to the original portrait image T using the initial positive expansion coefficient t1. RAW The degree of influence, generating the first animated image T A This is to prepare for subsequent adjustments, while enhancing identity features and improving controllability.

[0087] Please see Figure 1-2 Subtract the original character image from the first animated image to obtain the positive diffusion variable that reflects the changes in the image during the positive diffusion process;

[0088] Multiply the forward diffusion variable by the reverse expansion coefficient to obtain the adjustment base quantity;

[0089] The back diffusion adjustment is obtained by multiplying the base quantity by the correction coefficient matrix.

[0090] Subtract the back diffusion adjustment from the first animation image to obtain the second animation image;

[0091] The specific steps for obtaining the correction coefficient matrix are as follows:

[0092] Step 1: Extract single-region features from the local area of ​​the first animation image, including features of the eyes and mouth regions of the face;

[0093] Step 2: Compare the extracted single-region features with the expected features to obtain the feature error between the two;

[0094] Step 3: Based on the comparison results of all feature errors within the local region, generate the correction coefficient matrix for the local region.

[0095] In this embodiment, the calculation formulas for the second animation image and the correction coefficient matrix are as follows:

[0096] T B =T A -t2×(T A -T RAW )×XZ A ;

[0097]

[0098] in:

[0099] T B For the second animated image;

[0100] t2 is the reverse expansion coefficient;

[0101] XZ A This is the correction coefficient matrix;

[0102] (T A -T RAW The calculation result is a positive diffusion variable;

[0103] t2×(T A -T RAW The calculation result is used to adjust the base quantity;

[0104] t2×(T A -T RAW )×XZ A The calculation result is the reverse diffusion adjustment amount;

[0105] MSE is the correction coefficient matrix XZ A Feature error of a single region;

[0106] n represents the total amount of features in a single region;

[0107] f Ai f is the i-th single-region feature vector in the single-region features. 0i Let be the feature vector of the i-th desired region in the desired features.

[0108] This algorithm unit calculates the correction coefficient matrix XZ. A And combined with the reverse diffusion coefficient t2, for the first animation image T A With the original human image T RAW The differences were processed to obtain the adjusted second animation image T. B This is to correct local feature errors and improve image stability.

[0109] Please see Figure 1-2 The diffusion model training unit gradually adds Gaussian noise to the original human image on a computer to obtain the change in the second feature.

[0110] The second positive expansion coefficient is multiplied by the change in the second characteristic to obtain the second positive adjustment after the second positive adjustment;

[0111] The second positive adjustment is multiplied by the optimization coefficient matrix to obtain the overall adjustment after considering the overall image quality.

[0112] The second animation image is combined with the overall adjustment to obtain the third animation image.

[0113] In addition, the image acquisition module works in conjunction with the image expansion and extraction module, and uses the arithmetic mean method to obtain M in the second animation image. ’ A quadratic positive post-feature;

[0114] The image processing module uses the arithmetic mean method to obtain M-based results. ’ The average value of the quadratic features obtained by recombining the second feature with the quadratic positive backward features and the second feature weights.

[0115] The formula for calculating the third animation image in this embodiment is as follows:

[0116] T C =T B +t3×TB B ×Y B ;

[0117] in:

[0118] T c For the third animated image;

[0119] t3 is the second-order positive expansion coefficient;

[0120] TB B This is the change in the second characteristic;

[0121] Y B To optimize the coefficient matrix;

[0122] t3×TB B The calculation result is a second positive adjustment amount;

[0123] t3×TB B ×Y B The calculation result is the comprehensive adjustment amount.

[0124] The formula for calculating the mean of the second characteristic is as follows:

[0125]

[0126] in:

[0127] T ’ avg It is the average of the second characteristic;

[0128] M represents the total amount of the second positive feature;

[0129] T q For M ’ In a quadratic positive post-feature, the q-th feature, a q This refers to the q-th weight in the second feature weight reorganization;

[0130] The calculation result is the optimization coefficient matrix TB. B.

[0131] This algorithm unit utilizes the optimization coefficient matrix TB B And the second positive expansion coefficient t3, for the adjusted second animation image T B Further forward diffusion optimization is performed to generate a final, better, and more identity-enhanced third animated image T. c This is to improve the overall quality and coherence of the animation.

[0132] Example 2, please refer to Figure 1-2 Based on the comparison between the initial feature average and the first feature average benchmark, the value of the initial positive expansion coefficient is obtained. The specific diffusion value rules for the initial positive expansion coefficient are as follows:

[0133] If the initial feature average is greater than or equal to the first feature average benchmark, then the initial positive expansion coefficient is 0.7.

[0134] If the initial feature average is less than the first feature average benchmark, then the initial positive expansion coefficient is 0.3.

[0135] The specific diffusion value rules for the reverse expansion coefficient, based on the correction coefficient matrix and the benchmark correction coefficient matrix, are as follows:

[0136] If the correction coefficient matrix is ​​greater than the baseline correction coefficient matrix, then the reverse expansion coefficient is 0.6.

[0137] If the correction coefficient matrix is ​​smaller than the correction coefficient matrix base, then the reverse expansion coefficient is 0.4.

[0138] Based on the second feature average benchmark and the quadratic feature average, the specific diffusion value rules for the quadratic positive expansion coefficient are as follows:

[0139] If the average of the second feature is greater than or equal to the baseline of the second feature, then the value of the second positive expansion coefficient is 0.2.

[0140] If the average value of the second feature is less than the baseline value of the second feature, then the value of the second positive expansion coefficient is 0.3.

[0141] In this embodiment, the initial forward expansion coefficient t1 is used to control the first characteristic change TB during the forward diffusion process. RAW The degree of impact on the original image:

[0142] If the original image of the person is T RAW For images with high clarity and rich detail, the initial forward expansion coefficient t1 should be appropriately increased and adjusted to 0.7. This allows for more efficient use of the image's clarity features during the forward diffusion process, resulting in a higher quality first animation image T. ATo better preserve and enhance identity features, conversely, if the image is blurry and lacks details, the value of the initial positive expansion coefficient t1 should be reduced and adjusted to 0.3 to avoid excessive diffusion that could lead to image distortion.

[0143] In addition, when a person's identity features, including facial features and unique clothing, are very obvious, the value of the initial positive expansion coefficient t1 is increased to make these features more prominent in the diffusion process. If the identity features are not obvious, the value of the initial positive expansion coefficient t1 is decreased to prevent the diffusion process from introducing too many unnecessary changes.

[0144] The reverse diffusion coefficient t2 is used to control the reverse diffusion adjustment process (T) A -T RAW )×XZ A The calculation part is for the first animation image T A Degree of impact:

[0145] First animation image T A If the local features, including facial expressions and poses, have large errors compared to the desired features, the value of the back diffusion coefficient t2 is increased and adjusted to 0.6 to more effectively correct these errors. If the local feature errors are small, the value of the back diffusion coefficient t2 is decreased to avoid over-adjustment that could lead to overall image instability.

[0146] In addition, if the local feature error is uniformly distributed, the value of the back diffusion coefficient t2 can be appropriately increased to correct the error comprehensively. If the error is concentrated in certain specific areas, local adjustments can be made for these areas, while the overall back diffusion coefficient t2 can be appropriately reduced to maintain the overall stability of the image.

[0147] The second-order forward expansion coefficient t3 is used to control the feature increment during the second forward diffusion optimization process, i.e., TB. B ×Y B The extent to which the structural components affect the adjusted image:

[0148] If the adjusted second animation image T B High clarity and moderate contrast indicate good overall quality. The value of the second positive expansion coefficient t3 can be appropriately reduced and adjusted to 0.2 for minor optimization. If the image clarity is low and the contrast is poor, the value of the second positive expansion coefficient t3 should be increased to enhance the image features and quality.

[0149] In addition, when generating animation sequences, if it is necessary to ensure the continuity between adjacent frames, the value of the second positive expansion coefficient t3 can be appropriately increased, specifically adjusted to 0.4, so that the adjustment range of the image during the second positive diffusion process is greater and better adapts to the dynamic changes of the animation.

[0150] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for generating identity-enhanced character image animations based on a diffusion model, comprising an image acquisition module, an image expansion and extraction module, and an image processing module, characterized in that: The image expansion extraction module includes a diffusion model training unit and a feature extraction and evaluation unit; The specific generation method and steps are as follows: Step 1: Use the camera device of the image acquisition module to acquire the original image of the person and M original features of the original image of the person, and transmit them to the image processing module; Step 2: Using the diffusion model training unit, Gaussian noise is gradually added to the original human image on the computer to obtain the first feature change amount and transmit it to the image processing module. Step 3: Obtain from the image processing module the first feature average reference, the correction coefficient matrix reference, the second feature average reference, the weight coefficient group, and the diffusion coefficient group that sets the diffusion value rules. Step 4: The image processing module obtains the first animation image based on the original character image, the first feature change amount, and the diffusion coefficient group, and transmits it to the feature extraction and evaluation unit. Step 5: The feature extraction and evaluation unit evaluates the local feature error of the first animation image, obtains the correction coefficient matrix, and transmits it to the image processing module. Step 6: The image processing module obtains the second animation image based on the correction coefficient matrix, the first animation image, and the diffusion coefficient group, and transmits them to the image expansion and extraction module respectively. Step 7: Using the diffusion model training unit and the feature extraction and evaluation unit, obtain the second feature change and the optimization coefficient matrix respectively, and transmit them to the image processing module; Step 8: The image processing module obtains a third animation image based on the second animation image, the second feature change of the diffusion coefficient group, and the optimization coefficient matrix, and transmits it to the diffusion model training unit. Repeat steps 2-8 to complete the overall animation sequence.

2. The method for generating identity-enhanced character image animations based on a diffusion model according to claim 1, characterized in that: The diffusion coefficient set includes the primary forward diffusion coefficient, the reverse diffusion coefficient, and the secondary forward diffusion coefficient.

3. The method for generating identity-enhanced character image animations based on a diffusion model according to claim 2, characterized in that: The initial positive expansion coefficient is multiplied by the first feature change to obtain the initial positive adjustment amount; The first animated image is obtained by adding the initial positive adjustment amount to the original character image.

4. The method for generating identity-enhanced character image animations based on a diffusion model according to claim 3, characterized in that: The weight coefficient group includes a first feature weight reorganization and a second feature weight reorganization; The image processing module uses the arithmetic mean method to obtain the initial feature average value based on the M original features and the first feature weighting. Based on the comparison between the initial feature average and the first feature average benchmark, the value of the initial positive expansion coefficient is obtained. The specific diffusion value rules for the initial positive expansion coefficient are as follows: If the initial feature average value is greater than or equal to the first feature average benchmark, then the initial positive expansion coefficient is 0.

7. If the initial feature average value is less than the first feature average benchmark, then the initial positive expansion coefficient is 0.

3.

5. The method for generating identity-enhanced character image animations based on a diffusion model according to claim 4, characterized in that: Subtracting the original character image from the first animated image yields a positive diffusion variable. The forward diffusion variable is multiplied by the reverse expansion coefficient to obtain the adjustment base quantity; The adjustment base amount is multiplied by the correction coefficient matrix to obtain the back diffusion adjustment amount; The second animation image is obtained by subtracting the back diffusion adjustment amount from the first animation image; The specific steps for obtaining the correction coefficient matrix are as follows: Step 1: Extract single-region features from the local area of ​​the first animation image, including features of the eyes and mouth regions of the face; Step 2: Compare the extracted single-region features with the expected features to obtain the feature error between the two; Step 3: Generate the correction coefficient matrix of the local region based on the comparison results of all the feature errors in the local region.

6. The method for generating identity-enhanced character image animations based on a diffusion model according to claim 5, characterized in that: The specific diffusion value rules for the reverse expansion coefficient, based on the correction coefficient matrix and the reference of the correction coefficient matrix, are as follows: If the correction coefficient matrix is ​​greater than the correction coefficient matrix base, then the reverse expansion coefficient is 0.6; If the correction coefficient matrix is ​​smaller than the correction coefficient matrix base, then the reverse expansion coefficient is 0.

4.

7. The method for generating identity-enhanced character image animations based on a diffusion model according to claim 5, characterized in that: The diffusion model training unit gradually adds Gaussian noise to the original human image on a computer to obtain the second feature change. The second positive expansion coefficient is multiplied by the second characteristic change to obtain the second positive adjustment amount; The second positive adjustment amount is multiplied by the optimization coefficient matrix to obtain the comprehensive adjustment amount; The second animated image is combined with the aforementioned comprehensive adjustment amount to obtain the third animated image.

8. The method for generating identity-enhanced character image animations based on a diffusion model according to claim 7, characterized in that: The image acquisition module works in conjunction with the image expansion and extraction module, and uses the arithmetic mean method to obtain M in the second animation image. ’ A quadratic positive post-feature; The image processing module uses the arithmetic mean method to obtain the result based on M. ’ The average value of the quadratic features obtained by weighting the second feature and the quadratic positive backward features.

9. The method for generating identity-enhanced character image animations based on a diffusion model according to claim 8, characterized in that: Based on the second feature average benchmark and the quadratic feature average, the specific diffusion value rules for the quadratic positive expansion coefficient are as follows: If the average value of the quadratic features is greater than or equal to the second feature average benchmark, then the value of the quadratic positive expansion coefficient is 0.

2. If the average value of the quadratic features is less than the average benchmark of the second features, then the value of the quadratic positive expansion coefficient is 0.3.

Images