Three-dimensional style migration method and device and storage medium

Abstract

The invention discloses a three-dimensional style migration method and device and a storage medium, and relates to the technical field of computer vision, and the three-dimensional style migration method comprises the following steps: adjusting a preset stylized three-dimensional Gaussian sputtering model based on an obtained target intensity vector of an adjustable style adjuster, the style intensity vector of the adjustable style regulator is associated with the attribute of each Gaussian ball in the stylized three-dimensional Gaussian sputtering model, and the style intensity vector of the adjustable style regulator is associated with the attribute of each Gaussian ball in the stylized three-dimensional Gaussian sputtering model; and carrying out style migration on the target scene based on the adjusted stylized three-dimensional Gaussian sputtering model. According to the method, through the stylized three-dimensional Gaussian sputtering model with the adjustable strength, the style images with the ideal migration degree and multiple visual angles can be obtained, and the flexibility of migration degree control of three-dimensional style migration is improved.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to 3D style transfer methods, devices and storage media. Background Technology

[0002] With the deep integration of computer graphics and artificial intelligence technologies, 3D scene reconstruction and rendering technology has evolved from traditional geometric modeling to data-driven explicit representation.

[0003] Currently, the explicit nature of 3DGS (3D Gaussian Splatting) technology allows users to manipulate specific areas, enabling highly flexible stylized area control. However, its stylized output is relatively fixed, resulting in somewhat monotonous outputs. In practical applications, this can lead to two problems: firstly, over-stylization, causing a significant loss of geometric details and semantic content in the original scene (e.g., building textures being completely covered by abstract brushstrokes); secondly, insufficient stylization, resulting in weak style transfer in the 3D scene and an inability to create coherent color transitions and pattern blending in 3D space, thus failing to meet practical application requirements. Summary of the Invention

[0004] The main objective of this application is to provide a three-dimensional style transfer method, device, and storage medium, aiming to solve the technical problem of how to improve the flexibility of controlling the degree of style transfer in three-dimensional style transfer.

[0005] To achieve the above objectives, this application proposes a three-dimensional style transfer method, which includes:

[0006] Based on the target intensity vector of the adjustable style adjuster, the preset stylized 3D Gaussian sputtering model is adjusted to obtain the adjusted stylized 3D Gaussian sputtering model. The style intensity vector of the adjustable style adjuster is associated with the properties of each Gaussian sphere in the stylized 3D Gaussian sputtering model.

[0007] Style transfer is performed on the target scene based on the adjusted stylized 3D Gaussian sputtering model.

[0008] In one embodiment, before the step of adjusting the preset stylized 3D Gaussian sputtering model based on the target intensity vector of the acquired adjustable style adjuster, the method further includes:

[0009] Keeping the style intensity vector of the original style adjuster at a preset value, the original three-dimensional Gaussian sputtering model is trained based on preset training data to obtain the stylized three-dimensional Gaussian sputtering model.

[0010] Keeping the stylized 3D Gaussian sputtering model unchanged, the original style adjuster is trained based on the training data to obtain the adjustable style adjuster.

[0011] In one embodiment, the training data includes an original 3D scene and a stylized image, and the step of training the original style adjuster based on the training data includes:

[0012] Randomly sample the style adjustment parameters of the original style adjuster;

[0013] Based on the style adjustment parameters and the stylized 3D Gaussian sputtering model, the original 3D scene is rendered to obtain a first rendered image;

[0014] Based on the first rendered image and the stylized image, the style intensity vector corresponding to the style adjustment parameter is adjusted to obtain the adjusted style adjuster;

[0015] The adjusted style adjuster is redefined as the original style adjuster, and the step of randomly sampling the style adjustment parameters of the original style adjuster is performed until the number of random samplings reaches a preset sampling threshold.

[0016] In one embodiment, the step of adjusting the style intensity vector corresponding to the style adjustment parameter based on the first rendered image and the stylized image includes:

[0017] Calculate the full-style guided loss value between the first rendered image and the stylized image;

[0018] Calculate the zero-style guided loss value between the first rendered image and the original 3D scene;

[0019] The style intensity vector corresponding to the style adjustment parameter is adjusted based on the weighted sum between the full-style guided loss value and the zero-style guided loss value.

[0020] In one embodiment, the training data includes an original 3D scene and stylized images, and the step of training the original 3D Gaussian sputtering model based on the preset training data includes:

[0021] The original three-dimensional scene is rendered using the original three-dimensional Gaussian sputtering model to obtain a second rendered image;

[0022] Calculate the style loss value between the second rendered image and the stylized image;

[0023] Based on the style loss value, the original 3D Gaussian sputtering model is adjusted to obtain the trained 3D Gaussian sputtering model;

[0024] The trained 3D Gaussian sputtering model is redefined as the original 3D Gaussian sputtering model, and the step of rendering the original 3D scene using the original 3D Gaussian sputtering model is performed until the style loss value reaches a preset convergence threshold.

[0025] In one embodiment, before the step of training the original 3D Gaussian sputtering model based on preset training data, the method further includes:

[0026] The stylized image is obtained by performing style transfer on the input image using a two-dimensional diffusion model, wherein the input image is a multi-view image determined based on the original three-dimensional scene.

[0027] In one embodiment, the step of performing style transfer on the input image using a two-dimensional diffusion model includes:

[0028] An anchor view and a current view are determined from the input image, and the image features of the anchor view are projected onto a three-dimensional space based on the anchor viewpoint to obtain spatial features. The current view is any other view in the input image besides the anchor view.

[0029] The spatial features are projected onto the current viewpoint of each current view to obtain each distortion feature;

[0030] Each current view is updated based on the image features of the anchor view and the splicing features between each of the distortion features.

[0031] In one embodiment, prior to the step of re-identifying the trained 3D Gaussian sputtering model as the original 3D Gaussian sputtering model, the method further includes:

[0032] Based on the hit count, transparency, and occlusion relationship of each Gaussian sphere in the trained 3D Gaussian sputtering model, the rendering contribution value of each Gaussian sphere is determined.

[0033] Based on the rendering contribution values, the Gaussian spheres are filtered by a preset Gaussian filter to update the trained 3D Gaussian sputtering model.

[0034] Furthermore, to achieve the above objectives, this application also proposes a three-dimensional style transfer device, which includes:

[0035] The model building module is used to adjust the preset stylized 3D Gaussian sputtering model based on the target intensity vector of the obtained adjustable style adjuster, so as to obtain the adjusted stylized 3D Gaussian sputtering model. The style intensity vector of the adjustable style adjuster is related to the properties of each Gaussian sphere in the stylized 3D Gaussian sputtering model.

[0036] The style transfer module performs style transfer on the target scene based on the adjusted stylized 3D Gaussian sputtering model.

[0037] In addition, to achieve the above objectives, this application also proposes an electronic device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the three-dimensional style transfer method as described above.

[0038] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the three-dimensional style transfer method described above.

[0039] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the three-dimensional style transfer method as described above.

[0040] The one or more technical solutions proposed in this application have at least the following technical effects: First, by associating the style intensity vector of the adjustable style adjuster with the properties of each Gaussian sphere in the stylized 3D Gaussian sputtering model, the attribute offset of each Gaussian sphere can be flexibly adjusted through the style intensity vector, thereby controlling the degree of style transfer and providing a flexible and adjustable parameter basis for subsequent style transfer; Second, the stylized 3D Gaussian sputtering model is adjusted based on the obtained target intensity vector, and style transfer is performed on the target scene based on the adjusted stylized 3D Gaussian sputtering model. Due to the introduction of the target intensity vector, the degree of style transfer is no longer constant, but can be dynamically adjusted according to the user's wishes, thereby realizing flexible control of the degree of 3D style transfer. Attached Figure Description

[0041] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0042] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating an embodiment of the three-dimensional style transfer method of this application.

[0044] Figure 2A comparison chart of the effects of fixed output style transfer and intensity adjustable style transfer provided for this application;

[0045] Figure 3 A schematic diagram of the overall framework of the three-dimensional style transfer method provided in Embodiment 4 of this application;

[0046] Figure 4 This is a schematic diagram of the module structure of the three-dimensional style transfer device according to an embodiment of this application;

[0047] Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the 3D style transfer method in the embodiments of this application.

[0048] The purpose, features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0049] It should be understood that the specific embodiments described herein are merely used to explain the technical solutions of the present application and are not intended to limit the present application.

[0050] It should be noted that in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0051] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0052] Because existing 3D style transfer methods often use fixed outputs, it is difficult to achieve a controllable degree of style transfer, which may result in over-stylization or under-stylization. Please refer to [link / reference needed]. Figure 2 Figure (a) shows the effects of overstylization and understylization based on a fixed output. For example, in Figure (a), overstylization causes the background content of the original 3D scene to be completely lost, or understylization causes the 3D style transfer effect to be insignificant, making it difficult to meet the actual application requirements.

[0053] This application provides a solution. First, by associating the style intensity vector of an adjustable style adjuster with the properties of each Gaussian sphere in a stylized 3D Gaussian sputtering model, the attribute offset of each Gaussian sphere can be flexibly adjusted using the style intensity vector, thereby controlling the degree of style transfer and providing a flexible and adjustable parameter basis for subsequent style transfer. Second, before applying the model to style transfer, a two-stage training strategy is further adopted to achieve different degrees of style transfer. In the first stage, the style intensity vector of the original style adjuster is kept at a preset maximum value, and the original 3D Gaussian sputtering model is trained based on preset training data, enabling the 3D Gaussian sputtering model to learn... The first stage obtains stylized attribute information to create a stylized 3D Gaussian sputtering model, which can then be used for style transfer. In the second stage, the stylized 3D Gaussian sputtering model remains unchanged, and the original style adjuster is trained based on training data. This ensures the style adjuster can learn the style intensity vectors corresponding to different degrees of transfer, enhancing its adaptability to varying intensity requirements, thus obtaining an adjustable style adjuster. Finally, the stylized 3D Gaussian sputtering model is adjusted based on the target intensity vector obtained from the adjustable style adjuster, and style transfer is performed on the target scene based on the adjusted model. Please refer to [link to relevant documentation]. Figure 2 Figure (b) in the figure provides an effect diagram of intensity-adjustable stylization implemented based on the three-dimensional style transfer method of this application. The target intensity vector is determined by an adjustable style adjuster, and the attribute offset of each Gaussian sphere in the stylized three-dimensional Gaussian sputtering model is adjusted based on the target intensity vector, thereby achieving style transfer of any degree within the range and improving the flexibility of the transfer degree control of three-dimensional style transfer.

[0054] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions. The following description uses a terminal as the executing entity to illustrate this embodiment and the subsequent embodiments.

[0055] Based on this, embodiments of this application provide a three-dimensional style transfer method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the three-dimensional style transfer method of this application.

[0056] In this embodiment, the 3D style transfer method includes steps S10 to S20:

[0057] Step S10: Based on the target intensity vector of the adjustable style adjuster, the preset stylized three-dimensional Gaussian sputtering model is adjusted to obtain the adjusted stylized three-dimensional Gaussian sputtering model. The style intensity vector of the adjustable style adjuster is related to the properties of each Gaussian sphere in the stylized three-dimensional Gaussian sputtering model.

[0058] In one feasible embodiment, the properties of each Gaussian sphere in the stylized 3D Gaussian sputtering model are adjusted by the target intensity vector output by the adjustable style adjuster, thereby utilizing the ability of the 3D Gaussian sputtering model to render a 3D scene and achieving style transfer with an ideal degree of transfer.

[0059] Optionally, a 3D Gaussian sputtering model refers to a model structure composed of Gaussian spheres used to represent a 3D scene. Each Gaussian sphere is a multi-dimensional array containing the current position of a Gaussian sphere and its corresponding attributes such as color, transparency, and 3D covariance matrix. A style adjuster is a model component that outputs a style intensity vector. By controlling this vector, it controls the attribute offsets of each Gaussian sphere, thereby adjusting the degree of style transfer. The style intensity vector has the same dimension as the attribute dimension of the Gaussian spheres.

[0060] Optionally, the user can determine the degree of style transfer they wish to perform on the target scene by triggering the adjustable style adjuster; then, the terminal responds to the triggering operation on the adjustable style adjuster, determines the style transfer parameters, and further maps the style transfer parameters to the corresponding style adjustment vector, thereby controlling the offset of the attributes of each Gaussian sphere in the model, so as to control the degree of style transfer.

[0061] Optionally, a slider can be displayed in the corresponding front-end interface of the stylized 3D Gaussian sputtering model. This slider is an interactive display of an adjustable style adjuster, allowing the user to determine the degree of style transfer. Alternatively, it can be displayed in the user interface in the form of a slider, knob, etc. The user's trigger operation can be dragging, clicking, double-clicking, etc. This embodiment does not specifically limit this.

[0062] Optionally, before associating the style adjuster with the 3D Gaussian sputtering model, the style adjuster needs to be parameterized. This can be done by using a step function to quantize the sampling. For example, by using a preset step function, the continuous signal of the style adjuster can be converted into discrete values, and each discrete value can be encoded to associate it with the properties of the 3D Gaussian sputtering model.

[0063] Optionally, since the style intensity vector of the style adjuster can control the attribute offset of each Gaussian sphere, the attribute offset of each Gaussian sphere can be determined as a complete Gaussian neuron to realize the association between the style intensity vector and the attributes of each Gaussian sphere, thereby obtaining a stylized three-dimensional Gaussian sputtering model.

[0064] In one possible implementation, after step S10, the method further includes:

[0065] Step S11: Keep the style intensity vector of the original style adjuster at a preset value, and train the original three-dimensional Gaussian sputtering model based on the preset training data to obtain the stylized three-dimensional Gaussian sputtering model.

[0066] In one feasible embodiment, before training begins, the style intensity vector of the original style adjuster is initialized to 1 (preset value) and remains unchanged during training iterations to ensure that the original 3D Gaussian sputtering model is always optimized in the direction of maximizing style transfer during training, so that the properties of each Gaussian sphere in the model can gradually approach the target style.

[0067] Optionally, the style intensity vector can be kept at a preset maximum value (such as 110%) to ensure that the maximum degree of style transfer is only the target style shown in the training data, thereby avoiding over-stylization to some extent.

[0068] Optionally, the training data may include the original 3D scene and stylized images, wherein the original 3D scene serves as the input for model training and its output is multi-view rendered images; stylized images with the same viewpoint as each rendered image are determined from the training data, and the loss value between each rendered image and each stylized image is calculated, thereby back-optimizing the properties of each Gaussian sphere in the original 3D Gaussian sputtering model so that the original 3D Gaussian sputtering model learns the target style information contained in the stylized images.

[0069] Optionally, the training data may only include the original 3D scene. In the first iteration of training, since the model does not contain stylization information, the original 3D scene is processed by the original 3D Gaussian sputtering model in the adjustable 3D Gaussian sputtering model. The resulting multi-view initial rendering images also do not contain stylization information. The initial rendering images are then input into a preset 2D diffusion model, and the stylized images of each initial rendering image are output through the 2D diffusion model. The style loss value between each initial rendering image and each stylized image is calculated. Then, the original 3D Gaussian sputtering model is optimized in reverse based on the style loss value. The original 3D scene is reprocessed using the optimized original 3D Gaussian sputtering model, and the style loss value is calculated between each generated rendering image and the above stylized images. The above reverse optimization and subsequent steps are repeated until the style loss value converges or reaches a preset loss range.

[0070] It is understandable that by training the original Gaussian sputtering model, it can learn the target style features contained in the stylized image, and then style transfer can be achieved through the trained model.

[0071] Step S12: Keeping the stylized 3D Gaussian sputtering model unchanged, the original style adjuster is trained based on the training data to obtain an adjustable style adjuster.

[0072] In one feasible embodiment, the parameter space of the stylized 3D Gaussian sputtering model is frozen, and only the original style adjuster is optimized and trained. This enables the style adjuster to learn the stylized image's style transfer degree and the corresponding style intensity vector based on the training data and the reverse optimization process under different style transfer degrees. For example, it learns how much style adjustment vector needs to be adjusted for a 20% style transfer degree. Then, by adjusting the target intensity vector, the ideal transfer degree of style transfer can be achieved.

[0073] In this embodiment, the original 3D Gaussian sputtering model and the original style adjuster are trained and optimized through a two-stage training process, so that the 3D Gaussian sputtering model can achieve style transfer and the style adjuster can learn the changes in the degree of style transfer in order to achieve ideal control of the degree of transfer.

[0074] S20 performs style transfer on the target scene based on the adjusted stylized 3D Gaussian sputtering model.

[0075] In one feasible embodiment, the target scene is first input into a trained stylized 3D Gaussian sputtering model, mapping the 3D scene information contained in the target scene onto Gaussian spheres. Then, based on the obtained target intensity vector, the properties of each Gaussian sphere in the stylized 3D Gaussian sputtering model are adjusted to obtain an adjusted stylized 3D Gaussian sputtering model, thereby achieving the desired level of style transfer for the user. Subsequently, the target scene is processed through the adjusted stylized 3D Gaussian sputtering model to extract the 3D scene information contained in the target scene. Combined with the attribute offset information of the adjusted Gaussian spheres, a 2D image is regenerated, i.e., the style-transferred image, thus completing the style transfer of the target scene.

[0076] This embodiment provides a 3D style transfer method. By associating the style intensity vector of an adjustable style adjuster with the properties of each Gaussian sphere in a stylized 3D Gaussian sputtering model, the attribute offset of each Gaussian sphere can be flexibly adjusted using the style intensity vector, thereby controlling the degree of style transfer. Furthermore, through a two-stage training process, the stylized 3D Gaussian sputtering model learns the target style features, and the adjustable style adjuster learns standardized control over the degree of transfer. Subsequently, the stylized 3D Gaussian sputtering model is adjusted based on the obtained target intensity vector of the adjustable style adjuster, and style transfer is performed on the target scene based on the adjusted stylized 3D Gaussian sputtering model. Due to the introduction of the target intensity vector, the degree of style transfer is no longer static but can be dynamically adjusted according to the user's wishes, thus achieving flexible control over the degree of 3D style transfer.

[0077] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment can be referred to the above description, and will not be repeated hereafter. Based on this, the training data includes the original 3D scene and stylized images. Step S12, which involves training the original 3D Gaussian sputtering model based on the preset training data, includes:

[0078] Step E10: Render the original 3D scene using an adjustable 3D Gaussian sputtering model to obtain the second rendered image;

[0079] Step E20: Calculate the style loss value between the second rendered image and the stylized image;

[0080] Optionally, a stylized image refers to an image that has undergone two-dimensional style transfer, used to confirm the degree of stylization of the rendered image for training and learning of the three-dimensional Gaussian sputtering model.

[0081] Optionally, the style loss value refers to a quantitative indicator that measures the degree of difference between the style features of the rendered image and the style features of the stylized image. It can be L1 loss (mean absolute error loss), LPIPS (Learned Perceptual Image Patch Similarity Loss), or a weighted sum of loss values ​​calculated based on different loss functions. This embodiment does not specifically limit the loss function used.

[0082] Step E30: Adjust the original 3D Gaussian sputtering model according to the style loss value to obtain the trained 3D Gaussian sputtering model;

[0083] Optionally, based on the calculated style loss value, the gradients of each attribute of each Gaussian sphere in the original 3D Gaussian sputtering model are calculated layer by layer forward; then, the Gaussian sphere attributes, including position, covariance matrix, transparency and spherical harmonic coefficients, as well as a series of parameters such as the learning rate of the original 3D Gaussian sputtering model, are updated through optimization algorithms (such as stochastic gradient descent or Adam optimizer) to obtain the trained 3D Gaussian sputtering model.

[0084] Optionally, after updating the properties of the Gaussian spheres, Gaussian spheres with excessive transparency or low contribution to scene construction can be filtered out to improve the training efficiency of the model.

[0085] In one possible implementation, prior to step E40, the method further includes:

[0086] Step E31: Based on the number of hits, transparency, and occlusion relationship of each Gaussian sphere in the trained 3D Gaussian sputtering model, determine the rendering contribution value of each Gaussian sphere.

[0087] Optionally, each round of iterative training includes the generation of rendered images from different perspectives, and the number of hits refers to the number of times a Gaussian sphere is selected to construct a 3D scene during the image generation process from different perspectives; the occlusion relationship refers to the geometric coverage state of each Gaussian sphere, including but not limited to the number of times it is occluded and the occlusion ratio.

[0088] Optionally, the rendering contribution value is used to quantify the visual impact weight of each Gaussian sphere on the final rendering result. It can be obtained by weighting the different rendering contribution attributes of each Gaussian sphere. When the Gaussian sphere has fewer hits, lower transparency, or more occlusions, its corresponding rendering contribution value will be smaller.

[0089] For example, the rendering contribution attributes such as the number of hits, transparency, and occlusion relationship of each Gaussian sphere can be assigned weights, and their weighted sum can be determined as the rendering contribution value of each Gaussian sphere. During the training of the model, the weights of each rendering contribution attribute can be dynamically adjusted through a genetic algorithm in order to achieve a balance between rendering quality and the number of filters.

[0090] Step E32: Based on each rendering contribution value, filter each Gaussian sphere using a preset Gaussian filter to update the trained 3D Gaussian sputtering model.

[0091] Optionally, Gaussian spheres with rendering contribution values ​​lower than a preset contribution value threshold can be filtered using a Gaussian filter; or, the rendering contribution values ​​can be sorted and a preset number of Gaussian spheres can be filtered in order of contribution value from low to high.

[0092] Optionally, the Gaussian spheres used for filtering are marked as objects to be filtered. The covariance overlap between each object to be filtered and the Gaussian spheres retained in its neighborhood is calculated. Objects to be filtered with a covariance overlap lower than a preset threshold are retained to avoid excessive removal that could lead to loss of image information or scene holes. The neighborhood of each object to be filtered can be two or three times the standard range of the covariance matrix, or it can be a spherical region determined based on a distance threshold. This embodiment does not specifically limit this.

[0093] In this embodiment, by removing relatively unimportant Gaussian spheres, the model avoids assigning Gaussian neurons to redundant Gaussian spheres, thereby reducing the generation of artifacts in the stylized scene and improving the style transfer effect.

[0094] Step E40: The trained 3D Gaussian sputtering model is redefined as the original 3D Gaussian sputtering model, and the original 3D scene is rendered using the adjustable 3D Gaussian sputtering model until the style loss value reaches the preset convergence threshold.

[0095] Optionally, during the first stage of training, the style adjuster of the original style adjuster is kept at a preset value, and the original 3D Gaussian sputtering model associated with it is trained iteratively. During the iterative training, the original 3D scene is mapped to different perspectives through the original 3D Gaussian sputtering model to obtain a second rendered image, and the second rendered image is compared with the stylized image stored offline in advance. Alternatively, each second rendered image can be sequentially input into a 2D diffusion model to obtain a stylized image corresponding to each second rendered image, and then compared. Based on the style loss value between the second rendered image and the stylized image, the original 3D Gaussian sputtering model is adjusted, and the 3D scene is re-rendered through the adjusted 3D Gaussian sputtering model. The above process is repeated until the style loss value reaches the convergence threshold, and the trained stylized 3D Gaussian sputtering model is obtained. Then, 3D style transfer can be completed through the stylized 3D Gaussian sputtering model.

[0096] In this embodiment, through a reverse optimization training process based on style loss values, the three-dimensional Gaussian sputtering model gradually expresses the target style information contained in the stylized image, and thus style transfer can be achieved through the three-dimensional Gaussian sputtering model.

[0097] Based on the first and / or second embodiments of this application, in the third embodiment of this application, the content that is the same as or similar to the first and second embodiments described above can be referred to the above description and will not be repeated hereafter. Based on this, the training data includes the original 3D scene and stylized images. Step S13, which involves training the original style adjuster based on the training data, includes:

[0098] Step A10: Randomly sample the style adjustment parameters of the original style adjuster;

[0099] Optionally, style adjustment parameters are specific values ​​used to control style intensity and are an intuitive representation of the degree of style transfer. For example, the sliding ratio and rotation ratio of the style adjuster are displayed in the form of sliders, knobs, etc. Users can adjust the style intensity vector by controlling the style adjustment parameters, thereby controlling the offset of each Gaussian sphere in the stylized 3D Gaussian sputtering model, and thus controlling the degree of 3D style transfer.

[0100] Step A20: Based on the style adjustment parameters and the stylized 3D Gaussian sputtering model, perform style transfer on the original 3D scene to obtain the first rendered image;

[0101] Step A30: Based on the first rendered image and the stylized image, adjust the style intensity vector corresponding to the style adjustment parameters to obtain the adjusted style adjuster;

[0102] Optionally, a stylized image refers to an image that has undergone two-dimensional style transfer, used to determine the degree of stylization of the rendered image so that the style adjuster can be trained and learned.

[0103] In one feasible embodiment, the input of the original style adjuster is randomly sampled, i.e., the style adjustment parameters received are randomly determined, and the style adjustment parameters are mapped to the corresponding style intensity vector through the original style adjuster; then, the stylized 3D Gaussian sputtering model is adjusted according to the style intensity vector, and the original 3D scene is rendered through the adjusted 3D Gaussian sputtering model to obtain a first rendered image; then, the rendering degree of the first rendered image is determined according to the stylized image, and reverse optimization is performed according to the rendering degree to adjust the style intensity vector corresponding to the style adjustment parameter, so as to achieve the style transfer degree corresponding to the style adjustment parameter, and obtain the adjusted style adjuster.

[0104] In one feasible implementation, step A30, which involves adjusting the style intensity vector corresponding to the style adjustment parameters based on the first rendered image and the stylized image, includes:

[0105] Step A31: Calculate the full-style guided loss value between the first rendered image and the stylized image;

[0106] Step A32: Calculate the zero-style guided loss value between the first rendered image and the original 3D scene;

[0107] Optionally, the full-style guided loss value is used to measure the degree of difference between the rendered image generated by the model and the stylized image.

[0108] Optionally, the zero-style guided loss value is used to measure the differences in content and structure between the rendered image and the original 3D scene, making it convenient to use together with the full-style guided loss value to determine the current degree of style transfer.

[0109] Optionally, when calculating the full-style guided loss value, features such as color and texture features of the first rendered image and the stylized image are first extracted; then, the difference value (i.e., the loss value) between the two images is calculated by comparing each pixel or feature point using a preset loss function (such as mean squared error or perceptual loss function). A similar method can be used to calculate the zero-style loss value.

[0110] Step A33: Based on the weighted sum between the full-style guided loss value and the zero-style guided loss value, adjust the style intensity vector corresponding to the style adjustment parameter.

[0111] In one feasible embodiment, based on a preset weight allocation rule, a weighted sum between the full-style guided loss and the zero-style guided loss is calculated; then, using this weighted sum as the optimization objective, the loss value and gradient are calculated through the direction propagation algorithm, and the style intensity vector corresponding to the sampled style adjustment parameters is updated according to the gradient.

[0112] Optionally, the style adjustment parameter a can be determined as the weight parameter of the full style guidance loss value, and (1-a) can be determined as the weight parameter of the zero style guidance loss value; then, based on the above weight parameters, the weighted sum between the full style guidance loss and the zero style guidance loss is calculated; then, the weighted sum is used as the optimization objective to adjust the style intensity vector corresponding to the style adjustment parameter a.

[0113] For example, with the style adjustment parameter set to 20%, the weight of the full style guidance loss value is determined to be 20%, and the weight of the zero style guidance loss value is determined to be 80%. By calculating the weighted sum of the two, the style intensity vector of the style adjuster is further optimized so that the style intensity vector corresponds to the style adjustment parameter.

[0114] In this embodiment, the balance between full-style guidance and zero-style guidance is adjusted according to the style adjustment parameters, and the style intensity vector is adjusted according to the weighted sum of the loss values ​​of the two guidances, so that the style adjuster learns the correspondence between the style intensity vector and the degree of style transfer.

[0115] Step A40: Re-determine the adjusted style adjuster as the original style adjuster, and perform the step of randomly sampling the style adjustment parameters of the original style adjuster until the number of samplings reaches the preset sampling threshold.

[0116] Optionally, the sampling threshold is a conditional parameter for terminating iterative training of the original style adjuster. Experiments show that setting it to around 2000 iterations is sufficient to complete the training, resulting in a stylized 3D Gaussian sputtering model with superior performance.

[0117] In this embodiment, by randomly sampling style adjustment parameters, performing style transfer, and adjusting the vector based on the results, the accuracy of style transfer degree control is optimized while achieving adjustable transfer style.

[0118] Based on the first, second, and / or third embodiments of this application, in the fourth embodiment of this application, the content that is the same as or similar to the first, second, and third embodiments described above can be referred to the above description and will not be repeated hereafter. Furthermore, prior to the step of training the original three-dimensional Gaussian sputtering model based on preset training data in step S12, the following step is also included:

[0119] Step B10: Style transfer is performed on the input image using a two-dimensional diffusion model to obtain a stylized image, wherein the input image is a multi-view image determined based on the original three-dimensional scene.

[0120] In one feasible embodiment, a two-dimensional diffusion model can be used to inject the reference style features of the obtained reference style image into the input image, thereby completing the style transfer of the input image and obtaining a stylized image containing the reference style features.

[0121] Alternatively, in addition to the two-dimensional diffusion model, style injection and style alignment can also be achieved through an encoder-decoder approach.

[0122] Optionally, the multiple viewpoints in the input image are consistent with the viewpoints used to render the original 3D Gaussian scene through the 3D Gaussian sputtering model. As a result, the viewpoints of the subsequently obtained stylized images are consistent with the viewpoints of the rendered images, which facilitates the calculation of loss values.

[0123] In one feasible implementation, step B10, which involves style transfer of the input image using a two-dimensional diffusion model, includes:

[0124] Step B20: Determine the anchor view and the current view from the input image, and project the image features of the anchor view onto the three-dimensional space according to the anchor viewpoint to obtain spatial features. The current view is any other view in the input image besides the anchor view.

[0125] Optionally, the anchor view, which serves as a reference view for spatial feature projection, can be randomly selected by the system from the input image, or the user can specify a view with a wide field of view or high image quality as the anchor view; while the anchor viewpoint refers to the camera pose of the anchor view, that is, the shooting angle of the anchor view, including camera intrinsic and extrinsic parameters.

[0126] Optionally, during the 3D projection process, the image coordinates contained in the 2D image are first converted into normalized planar coordinates, then combined with the predicted depth to convert them into 3D points in the camera coordinate system, and finally converted into 3D points in the world coordinate system using the camera pose, so as to match and stitch them with the image features of the current viewpoint in the current view. The preset depth is a 2D array containing the distance from each pixel in the 2D image to the camera, which can be obtained through a depth sensor, stereo matching, or depth estimation network.

[0127] Step B30: Project the spatial features onto the current viewpoint of each current view to obtain each distortion feature; Optionally, the above step of reprojecting the three-dimensional spatial features onto the current viewpoint to obtain the distortion content features (distortion features) includes: First, determining the current viewpoint of the current view, i.e., the camera pose of the current view, including the camera intrinsic and extrinsic parameters (rotation matrix and translation vector of the camera in the world coordinate system); Then, using the predicted depth between the current viewpoint and the camera, determining the three-dimensional spatial feature points, and using the camera's rotation matrix and translation vector, transforming the three-dimensional spatial feature points to the camera coordinate system; Then, using the camera's intrinsic parameters for projection to obtain the distortion features of the current view.

[0128] Step B40: Update each current view based on the image features of the anchor view and the splicing features between each distortion feature.

[0129] Optionally, in a two-dimensional diffusion model, the above features can be concatenated and updated using the key, value, and query features of the self-attention layer.

[0130] For example, on the one hand, Gaussian noise (related information from the style image) is gradually added to the input image through a two-dimensional diffusion model to generate a stylized image; on the other hand, image features of the input image or stylized image are extracted through a self-attention layer, and the distorted features of the current view A are obtained through the aforementioned feature warping method; then, the image features of the anchor view and the distorted features are concatenated to obtain concatenated features, which are used as the "key" and "value" features of the anchor view. These features are then matched and updated with the "key" and "value" features of the current view A through mutual self-attention, thereby ensuring style consistency between the current view A and the anchor view. In addition to "key" and "value" features, "query" features can also be used for matching and updating.

[0131] Optionally, the above-mentioned process of feature stitching and feature updating can be performed before or after generating a stylized image from a two-dimensional model, or simultaneously. This embodiment does not specifically limit this.

[0132] In this embodiment, cross-view style alignment is achieved by stitching features between views from different perspectives, further improving the stability of content calibration. The resulting stylized image has multi-view characteristics. Figure 1 Consistency is beneficial for improving the training effect of stylized 3D Gaussian sputtering models.

[0133] For example, to help understand the implementation flow of the 3D style transfer method obtained by combining the above embodiments one, two, and three, please refer to... Figure 3 , Figure 3 A schematic diagram of the overall framework of a 3D style transfer method is provided, specifically:

[0134] Figure 3 The original content scene can be stylized to an ideal degree by using an intensity-adjustable style injector.

[0135] Figure 3 A schematic diagram of a training process for an intensity-adjustable style injector (including an adjustable style adjuster and a stylized 3D Gaussian sputtering model) is provided. Figure 3 In the model, Δμ, ΔR, ..., Δσ, and Δc represent the attribute offsets of the Gaussian spheres, such as position offset, rotation offset, size offset, and color offset. First, the attributes of each Gaussian sphere in the original 3D Gaussian sputtering model are associated with the style intensity vector of the style adjuster, forming Gaussian neurons to control the attribute offsets such as position offset, rotation offset, size offset, and color offset. Then, these Gaussian neurons are combined with other parameters in the original 3D Gaussian model to obtain the stylized 3D Gaussian sputtering model. Second, the two-stage training process of the intensity-adjustable style injector involves an adjustable stylization guide. For example, during the first iteration of training the original 3D Gaussian sputtering model, the output rendered image is input into the adjustable stylization guide to obtain the corresponding stylized image. The loss value is calculated based on the difference between the rendered image and the stylized image, and backpropagation is performed to update each Gaussian sphere and / or the style adjuster until the loss value reaches a preset loss condition, at which point the iterative training ends, resulting in the stylized 3D Gaussian sputtering model. In addition, before training, a stylized image corresponding to the input image can be generated by an adjustable stylization guide, wherein the input image is a multi-view image determined based on the original 3D scene.

[0136] Figure 3 It also includes a schematic diagram of generating stylized images using an adjustable stylization guide (i.e., a two-dimensional diffusion model with an added self-attention layer). On one hand, the stylistic features of a reference stylization image are injected into the input image (including the anchor view and the current view) through the two-dimensional diffusion model; on the other hand, style alignment between input images from different viewpoints is achieved through mutual self-attention of features such as "key, value, and query" in the self-attention layer. After the above two aspects of image processing, the stylized result of the input image is obtained. At the same time, the input image can be used to provide zero-style guidance for training the style adjuster, and the stylized result can be used to provide full-style guidance for training the style adjuster.

[0137] It is important to note that the adjustable stylization guide is already trained before training the stylized 3DGS model. Therefore, during the training of the stylized 3DGS model, the adjustable stylization guide is frozen and its parameters do not change.

[0138] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the three-dimensional style transfer method of this application. Any simple transformations based on this technical concept are within the protection scope of this application.

[0139] This application also provides a three-dimensional style transfer device. Please refer to... Figure 4 The three-dimensional style transfer device includes:

[0140] The model building module 10 is used to adjust the preset stylized three-dimensional Gaussian sputtering model based on the target intensity vector of the obtained adjustable style adjuster, so as to obtain the adjusted stylized three-dimensional Gaussian sputtering model. The style intensity vector of the adjustable style adjuster is related to the properties of each Gaussian sphere in the stylized three-dimensional Gaussian sputtering model.

[0141] Style transfer module 20 is used to perform style transfer on the target scene based on the adjusted stylized 3D Gaussian sputtering model.

[0142] The 3D style transfer apparatus provided in this application, employing the 3D style transfer method described in the above embodiments, can solve the technical problem of how to improve the flexibility of controlling the degree of 3D style transfer. Compared with the prior art, the beneficial effects of the 3D style transfer apparatus provided in this application are the same as those of the 3D style transfer method provided in the above embodiments, and other technical features in the 3D style transfer apparatus are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0143] This application provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the three-dimensional style transfer method in the first embodiment described above.

[0144] The following is for reference. Figure 5 The diagram illustrates a structural schematic of an electronic device suitable for implementing embodiments of this application. The electronic devices in these embodiments may include, but are not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0145] like Figure 5 As shown, the electronic device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the electronic device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. The communication device 1009 allows the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although the diagrams show electronic devices with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented alternatively.

[0146] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0147] The electronic device provided in this application, employing the 3D style transfer method described in the above embodiments, can solve the technical problem of how to improve the flexibility of controlling the degree of 3D style transfer. Compared with the prior art, the beneficial effects of the electronic device provided in this application are the same as those of the 3D style transfer method provided in the above embodiments, and other technical features of the electronic device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0148] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0149] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0150] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the three-dimensional style transfer method in the above embodiments.

[0151] The computer-readable storage medium provided in this application embodiment may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0152] The aforementioned computer-readable storage medium may be included in an electronic device or may exist independently without being assembled into an electronic device.

[0153] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by an electronic device, the electronic device constructs a stylized 3D Gaussian sputtering model. The stylized 3D Gaussian sputtering model includes: adjusting a preset stylized 3D Gaussian sputtering model based on the target intensity vector of an acquired adjustable style adjuster to obtain an adjusted stylized 3D Gaussian sputtering model, wherein the style intensity vector of the adjustable style adjuster is associated with the properties of each Gaussian sphere in the stylized 3D Gaussian sputtering model; and performing style transfer on a target scene based on the adjusted stylized 3D Gaussian sputtering model.

[0154] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0155] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a combination of dedicated hardware and computer instructions.

[0156] The modules described in the embodiments of the present application may be implemented in software or hardware, wherein the name of a module does not necessarily limit the unit itself.

[0157] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described 3D style transfer method. This solves the technical problem of how to improve the flexibility of controlling the degree of 3D style transfer. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the 3D style transfer method provided in the above embodiments, and will not be repeated here.

[0158] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the three-dimensional style transfer method described above.

[0159] The computer program product provided in this application can solve the technical problem of how to improve the flexibility of controlling the degree of style transfer in 3D style transfer. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the 3D style transfer method provided in the above embodiments, and will not be repeated here.

[0160] The above description is only part of the embodiments of the present application and does not limit the patent scope of the present application. All equivalent structural transformations made by using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A three-dimensional style transfer method, characterized in that, The 3D style transfer method includes: Based on the target intensity vector of the adjustable style adjuster, the preset stylized 3D Gaussian sputtering model is adjusted to obtain the adjusted stylized 3D Gaussian sputtering model. The style intensity vector of the adjustable style adjuster is associated with the properties of each Gaussian sphere in the stylized 3D Gaussian sputtering model. Style transfer is performed on the target scene based on the adjusted stylized 3D Gaussian sputtering model.

2. The three-dimensional style transfer method as described in claim 1, characterized in that, Before the step of adjusting the preset stylized 3D Gaussian sputtering model based on the target intensity vector of the acquired adjustable style adjuster, the method further includes: Keeping the style intensity vector of the original style adjuster at a preset value, the original three-dimensional Gaussian sputtering model is trained based on preset training data to obtain the stylized three-dimensional Gaussian sputtering model. Keeping the stylized 3D Gaussian sputtering model unchanged, the original style adjuster is trained based on the training data to obtain the adjustable style adjuster.

3. The three-dimensional style transfer method as described in claim 2, characterized in that, The training data includes the original 3D scene and stylized images. The step of training the original style adjuster based on the training data includes: Randomly sample the style adjustment parameters of the original style adjuster; Based on the style adjustment parameters and the stylized 3D Gaussian sputtering model, the original 3D scene is rendered to obtain a first rendered image; Based on the first rendered image and the stylized image, the style intensity vector corresponding to the style adjustment parameter is adjusted to obtain the adjusted style adjuster; The adjusted style adjuster is redefined as the original style adjuster, and the step of randomly sampling the style adjustment parameters of the original style adjuster is performed until the number of random samplings reaches a preset sampling threshold.

4. The three-dimensional style transfer method as described in claim 3, characterized in that, The step of adjusting the style intensity vector corresponding to the style adjustment parameter based on the first rendered image and the stylized image includes: Calculate the full-style guided loss value between the first rendered image and the stylized image; Calculate the zero-style guided loss value between the first rendered image and the original 3D scene; The style intensity vector corresponding to the style adjustment parameter is adjusted based on the weighted sum between the full-style guided loss value and the zero-style guided loss value.

5. The three-dimensional style transfer method as described in claim 2, characterized in that, The training data includes the original 3D scene and stylized images. The step of training the original 3D Gaussian sputtering model based on the preset training data includes: The original three-dimensional scene is rendered using the original three-dimensional Gaussian sputtering model to obtain a second rendered image; Calculate the style loss value between the second rendered image and the stylized image; Based on the style loss value, the original 3D Gaussian sputtering model is adjusted to obtain the trained 3D Gaussian sputtering model; The trained 3D Gaussian sputtering model is redefined as the original 3D Gaussian sputtering model, and the step of rendering the original 3D scene using the original 3D Gaussian sputtering model is performed until the style loss value reaches a preset convergence threshold.

6. The three-dimensional style transfer method as described in claim 3 or 5, characterized in that, Before the step of training the original 3D Gaussian sputtering model based on preset training data, the method further includes: The stylized image is obtained by performing style transfer on the input image using a two-dimensional diffusion model, wherein the input image is a multi-view image determined based on the original three-dimensional scene.

7. The three-dimensional style transfer method as described in claim 6, characterized in that, The steps of performing style transfer on the input image using a two-dimensional diffusion model include: An anchor view and a current view are determined from the input image, and the image features of the anchor view are projected onto a three-dimensional space based on the anchor viewpoint to obtain spatial features. The current view is any other view in the input image besides the anchor view. The spatial features are projected onto the current viewpoint of each current view to obtain each distortion feature; Each current view is updated based on the image features of the anchor view and the splicing features between each of the distortion features.

8. The three-dimensional style transfer method as described in claim 5, characterized in that, Before the step of re-identifying the trained 3D Gaussian sputtering model as the original 3D Gaussian sputtering model, the method further includes: Based on the hit count, transparency, and occlusion relationship of each Gaussian sphere in the trained 3D Gaussian sputtering model, the rendering contribution value of each Gaussian sphere is determined. Based on the rendering contribution values, the Gaussian spheres are filtered by a preset Gaussian filter to update the trained 3D Gaussian sputtering model.

9. An electronic device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the three-dimensional style transfer method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the three-dimensional style transfer method as described in any one of claims 1 to 8.

Images