Three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation

Abstract

The invention discloses a three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation, and the method comprises the steps: firstly generating a geometric structure and surface texture of a three-dimensional packaging model through Gaussian distribution, and obtaining particle distribution; secondly, generating a particle attribute data set containing physical consistency parameters on the basis of particle distribution in combination with a physical simulation technology based on a particle system, globally optimizing the three-dimensional model by adopting a ray tracing and real-time rendering technology, and then combining rendered three-dimensional model features with multi-modal data input by a user to obtain a multi-modal data set; and generating an adaptive design template in combination with a text collaborative prompt learning algorithm. And finally, integrating the diffusion model and a neural radiation field NeRF technology, and generating a detailed three-dimensional packaging model based on physical consistency parameters and a modeling result in combination with a design template. According to the method, the user requirements can be accurately analyzed, accurate and efficient personalized customization is realized, and the user experience is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of computer graphics and three-dimensional modeling, and in particular to a three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation. Background Art

[0002] Traditional 3D packaging modeling methods mainly rely on parametric modeling tools (such as Maya and Blender) and 3D scanning and reconstruction technology. In parametric modeling, designers need to manually define the geometric topology and add texture maps layer by layer. This process is time-consuming and has poor adaptability to complex surfaces (such as special-shaped bottles and wrinkled packaging). Although scanning and reconstruction-based technologies can quickly obtain point clouds on the surface of objects, it is difficult to generate customized textures that meet the design intent, and the light field characteristics of translucent materials (such as glass and plastic) are not accurately restored. In particular, when using traditional particle systems to generate surface textures, due to the lack of a dynamic response mechanism to changes in curvature, particles are prone to excessive accumulation or sparse loss in concave and convex areas, resulting in distortion of texture details (such as broken gradient patterns and blurred relief effects).

[0003] Existing methods suffer from significant deficiencies in physical consistency and interactivity. First, the texture generation process fails to incorporate physics engine simulation, relying solely on random distribution algorithms for particle motion paths. The effects of environmental factors like gravity and collisions on texture morphology are ignored, resulting in light and shadow reflections and material properties (such as metallic sheen and matte texture) that deviate from real-world physical laws. Second, design tools generally employ offline rendering, requiring users to wait minutes for rendering to complete after modifying model parameters before viewing the results. Furthermore, dimensional adjustments are limited to basic properties like geometric dimensions, preventing real-time modification of higher-level parameters like texture density (e.g., label patchiness) and dynamic lighting (e.g., ambient occlusion intensity), leading to inefficient design iterations. Furthermore, existing systems lack the ability to integrate multimodal inputs. Designer-provided text descriptions (e.g., "vintage leather texture") or reference images are difficult to accurately map to the 3D model space through feature decoupling and cross-modal alignment techniques. Consequently, generated results often exhibit semantic ambiguity (e.g., misinterpreting "matte" as "matte") or style conflicts.

[0004] Although Neural Radiance Field (NeRF) technology has achieved high-precision 3D reconstruction through implicit representation in recent years, its static modeling characteristics cannot support dynamic physical interactions, and training takes hours, making it difficult to meet the real-time iteration requirements of packaging design. Diffusion models, while excellent for generating 2D images, suffer from cross-viewpoint texture inconsistencies (e.g., a break between the front and side patterns of a bottle) and low detail resolution when directly applied to 3D scenes. These limitations severely hinder the digital advancement of the packaging design industry, necessitating the urgent need for innovative technical solutions. Summary of the Invention

[0005] In response to the above problems, the present invention provides a three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation, which combines physical simulation, ray tracing, multimodal fusion and deep learning technology to achieve highly realistic and interactive packaging design, alleviate the efficiency reduction caused by the inability to quickly generate three-dimensional modeling data, and solve the problem of not being able to meet personalized customization needs.

[0006] The technical solution adopted by the present invention to solve the technical problem is to provide a three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation, comprising the following steps:

[0007] Step 1: Use the Gaussian distribution model to generate the geometric structure and surface texture of the three-dimensional packaging model. The distribution density and range of particles in space (i.e., discrete three-dimensional basic units generated based on the Gaussian sputtering algorithm, carrying attributes such as position, color, and energy) are controlled by the mean, standard deviation, and sputtering angle. Combined with dynamic parameter adjustment technology, the sputtering of particles on complex packaging forms is made more uniform and in line with natural distribution laws.

[0008] Step 2: Based on the particle distribution generated in step 1, combined with the physical simulation technology based on the particle system, the particle acceleration, direction change, collision and energy attenuation process are dynamically calculated to generate a particle attribute data set containing physical consistency parameters (such as color gradient, transparency change, collision traces) to ensure high-fidelity visual effects of the texture.

[0009] Step 3. Based on the particle attribute dataset output in Step 2, use ray tracing and real-time rendering technology to globally optimize the 3D model: simulate complex light and shadow interactions through particle color and transparency data, enhance material surface details (such as frosted particles and embossed bumps) by combining particle collision traces, and use energy attenuation parameters to generate dynamic transparency transition effects, thereby improving the realism and detail expression of the packaging design.

[0010] Step 4: Combine the 3D model features rendered in step 3 with the multimodal data (text, image) input by the user. Combined with the text collaborative prompt learning algorithm, the text description input by the user is parsed into a high-dimensional semantic vector. The rendering results (such as lighting style and texture density) are aligned with user needs using cross-modal matching technology to generate an adaptive design template.

[0011] Step 5: Integrate the diffusion model and NeRF (neural radiant field) technology. Based on the physical consistency parameters of step 2 and the modeling results of step 3, combined with the design template, a highly detailed 3D packaging model is generated. The diffusion model uses particle color and energy data to generate micro-textures (such as metal oxide layers and printed patterns) in stages, while NeRF reconstructs transparent materials and reflective properties that are consistent from multiple perspectives based on the rendered light field.

[0012] Step 6. Support real-time adjustment of generated results: Users can dynamically modify the particle distribution parameters of step 1, the physical simulation rules of step 2, the light and shadow configuration of step 3, or the texture generation weight of step 5 through the interactive interface. Based on the dependencies of the above steps, the system synchronously updates the model geometry, texture style, and light field environment through a real-time feedback mechanism, and ultimately outputs a 3D packaging modeling result that meets user needs, completing a complete closed-loop process from parametric design to highly realistic packaging models.

[0013] In step 1, based on the three-dimensional Gaussian sputtering algorithm, by dynamically adjusting the mean, standard deviation, sputtering angle and curvature adaptive technology, uniform particle distribution of complex packaging shapes is achieved, and the geometric and texture foundation of a highly realistic three-dimensional packaging model is constructed. The specific implementation is as follows:

[0014] By adjusting the mean and standard deviation of the Gaussian distribution, the sputtering particle distribution that conforms to the packaging morphology characteristics is dynamically generated, and the spatial mapping of the particle position is performed in combination with three-dimensional voxel grid technology to ensure the sputtering uniformity of particles on surfaces with different curvatures.

[0015] Furthermore, in step 2, the physical simulation technology based on the particle system is combined to dynamically calculate the acceleration, direction change, collision and energy attenuation process of the particles, and generate a particle attribute data set containing physical consistency parameters (such as color gradient, transparency change, collision traces) to ensure high-fidelity visual effects of the texture; the specific implementation is:

[0016] (1) By introducing environmental parameters such as gravity and air resistance, the acceleration changes of particles in three-dimensional space are simulated.

[0017] (2) Use the collision detection algorithm to simulate the collision behavior between particles and the model surface and calculate the particle distribution path after sputtering.

[0018] (3) According to the energy attenuation law of particle motion, the particle color and transparency are adjusted to generate a more realistic texture effect.

[0019] Furthermore, in step 3, based on the particle attribute dataset from step 2, ray tracing and real-time rendering technology are used to globally optimize the 3D model. Light and shadow interactions and material details (such as frosted particles and embossed bumps) are simulated through particle color, transparency, and collision traces. Energy attenuation parameters are used to generate dynamic transparency transitions, significantly enhancing the realism and detail of the packaging design. Specifically, this is achieved as follows:

[0020] (1) Based on the BRDF (bidirectional reflectance distribution function) model, the light reflection characteristics of different materials are modeled, and a mathematical relationship between the material surface reflectivity, roughness and incident / outgoing angle is established to provide a physically accurate reflection calculation basis for ray tracing.

[0021] (2) Based on the BRDF modeling results of step (1), the light is divided into primary light (direct illumination), diffuse light (indirect diffuse reflection) and secondary reflection light (specular reflection / refraction) through light layering technology, and an independent rendering strategy is designed for each layer of light. The specific rendering strategy for each light is as follows:

[0022] Main ray: Calculates specular reflection and Fresnel effect directly in combination with BRDF parameters.

[0023] Diffuse Light: Optimizes sampling density based on material roughness parameters to improve noise suppression efficiency in diffuse areas.

[0024] Secondary reflected rays: For reflected rays, computing resources are dynamically allocated based on the BRDF specular weight, giving priority to multiple reflection paths in highlight areas.

[0025] (3) Combining the light reflection characteristics of different materials in step (1) with the light stratification results in step (2), a transparency processing algorithm based on volume rendering is adopted to simulate the light transmission and internal scattering effects of transparent materials (such as glass and plastic) through the ray absorption-scattering model, and the light path integral is optimized according to the BRDF refractive index parameters, ultimately achieving the visual fusion of multi-layer refraction and surface reflection of transparent materials, thereby improving the realism of complex packaging models.

[0026] Furthermore, in step 4, the 3D model features rendered in step 3 are combined with the user's multimodal input (text, image), the semantic vector is parsed through the text collaborative prompt learning algorithm, and the rendering results (such as lighting style and texture density) are aligned with the user's needs using cross-modal matching technology to generate an adapted 3D packaging design template. The specific implementation is as follows:

[0027] A multi-layer neural network is used to process the text descriptions and image content input by the user, and cross-modal matching technology is used to automatically screen image features related to the design goals. In combination with a gradient descent-based optimization algorithm, a design template that meets diversity and aesthetic requirements is generated.

[0028] Specifically, in step 5, the diffusion model and NeRF technology are integrated. Based on the physical consistency parameters of step 2 and the light and shadow simulation results of step 3, the diffusion model generates micro-textures (such as metal oxide layers and printed patterns) in stages. NeRF (neural radiant field) reconstructs transparent materials and reflective properties that are consistent across multiple perspectives, ultimately generating a highly detailed 3D packaging model. The specific implementation is as follows:

[0029] (1) Using the staged generation strategy of the diffusion model, a rough texture layout is first generated, and then the texture details are gradually improved through multi-step denoising to obtain a refined texture layout.

[0030] (2) For the refined texture layout, the light field is reconstructed using multi-resolution sampling in combination with Mip-NeRF (multi-scale representation of neural radiance field) technology to ensure the texture consistency of the model under multi-angle observation.

[0031] (3) The accuracy of texture generation and light field reconstruction is optimized through a joint loss function to improve the clarity and realism of the 3D packaging model.

[0032] Furthermore, in step 6, the user is supported to adjust the parameters of each step (particle distribution, physical rules, light and shadow configuration, texture weight) in real time through the interactive interface. The system synchronously updates the model attributes in real time based on the dependency relationship, and ultimately generates a highly realistic 3D packaging modeling result, realizing a complete closed-loop process from parametric design to finished product output. The specific implementation is as follows:

[0033] The interactive interface provides users with flexible operation tools by combining multiple methods such as clicking, dragging, and numerical input. It supports real-time preview of the specific impact of different adjustment parameters on the design effect, and saves the adjustment path through the intelligent history recording function for subsequent review or improvement.

[0034] The technical solutions provided in the embodiments of the present invention have at least the following technical effects:

[0035] Using three-dimensional Gaussian sputtering technology, three-dimensional packaging models can be quickly generated, significantly improving modeling efficiency; by combining the text description and image content entered by the user, as well as cross-modal matching technology, user needs can be accurately analyzed, accurate and efficient personalized customization can be achieved, and the user experience can be enhanced.

[0036] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of an embodiment of the present invention;

[0038] Figure 2 is a flow chart of a three-dimensional Gaussian sputtering algorithm according to an embodiment of the present invention;

[0039] Figure 3 is a multimodal schematic diagram of an embodiment of the present invention;

[0040] Figure 4 It is a framework diagram of the integration of diffusion model and neural radiation field NeRF technology in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] An embodiment of the present invention relates to a three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation, which uses Gaussian sputtering technology to generate the geometric structure and surface details of the three-dimensional model, and uses a multimodal technology-based fusion text collaborative prompt learning and three-dimensional reconstruction algorithm to address the problem of personalized customization. According to the text description or image information input by the user, a packaging design scheme that meets the requirements is quickly generated, and then the diffusion model and NeRF (neural radiation field) technology are integrated to generate a highly detailed three-dimensional packaging model.

[0042] like Figure 1 As shown, the following steps are included:

[0043] Step 1: Use a Gaussian distribution model to generate the geometric structure and surface texture of a 3D packaging model. Control the density and range of particle distribution in space through the mean, standard deviation, and sputtering angle. Combined with dynamic parameter adjustment technology, this ensures that particle sputtering on complex packaging forms is more uniform and conforms to natural distribution patterns.

[0044] The step one includes the following steps:

[0045] Step 101: Define the mean (μ) and standard deviation (σ) of the Gaussian distribution to control the distribution center and diffusion range of particles in three-dimensional space.

[0046] refer to Figure 2 , we can see that the process of Gaussian sputtering algorithm includes:

[0047] Step 102: Dynamically generate a sputtering particle distribution that conforms to the packaging morphology by adjusting the mean and standard deviation of the Gaussian distribution. Combined with three-dimensional voxel grid technology, perform spatial mapping of particle positions to ensure uniform sputtering of particles on surfaces with different curvatures. Specifically, the process includes:

[0048] S1021. Dynamically adjust the mean and standard deviation based on the geometric characteristics and texture requirements of the packaging model. For example, for a raised area, the mean may be biased toward the center of the area, and the standard deviation may be adjusted based on the size of the area.

[0049] S1022. Use a Gaussian distribution model to randomly generate the position coordinates (x, y, z) of the particles in three-dimensional space, and then limit the distribution range of the particles according to the sputtering angle to ensure that the particles are sputtered within a specific angle range to avoid excessive deviation from the target surface.

[0050] S1023. Divide the three-dimensional space into a uniform voxel grid, where each voxel represents a small spatial unit. Then, map the generated particle positions to the corresponding voxels, calculate the particle density within each voxel, and adjust the particle density within the voxel according to the curvature distribution:

[0051] For areas of high curvature, increase the density of particles;

[0052] For low curvature areas, reduce the density of particles;

[0053] S1024. Monitor the distribution of particles in the voxel grid in real time, calculate the uniformity index of the particle distribution, and then dynamically adjust the mean and standard deviation of the Gaussian distribution based on the uniformity results to optimize the sputtering effect of particles on complex packaging forms.

[0054] S1025. Mapping the adjusted particle distribution onto the surface of the packaging model to generate a texture distribution that conforms to natural laws, and then using an interpolation or smoothing algorithm to ensure a natural connection of the texture in the transition area to avoid mutations or discontinuities.

[0055] When using, combine the contents in steps 101 to 102:

[0056] The mean (μ) and standard deviation (σ) of the Gaussian distribution are defined to control the distribution center point and diffusion range of particles in three-dimensional space, and the distribution center is precisely located. By adjusting the mean and standard deviation of the Gaussian distribution, the sputtering particle distribution that conforms to the packaging morphological characteristics is dynamically generated, and the three-dimensional voxel grid technology is combined to perform spatial mapping of the particle position to ensure the sputtering uniformity of particles on surfaces with different curvatures, enhance the naturalness of the texture and the adaptability of details, and avoid discontinuities caused by morphological differences.

[0057] Step 2: Combined with particle system-based physical simulation technology, dynamically calculate the particle acceleration, direction change, collision and energy decay process to ensure that the generated texture has physical consistency and high-fidelity visual effects;

[0058] The step 2 includes the following steps:

[0059] Step 201: According to Newton's second law, gravity acceleration (g) and air resistance coefficient (k) are introduced to dynamically calculate the acceleration change of the particle in three-dimensional space:

[0060] a=gk·v

[0061] Where a is the instantaneous acceleration of the particle and v is the current velocity of the particle. The velocity and position of the particle are updated through numerical integration (such as the Euler method or the Verlet algorithm) to simulate the natural motion trajectory under the influence of environmental forces.

[0062] Step 202: Use the AABB / OBB bounding box collision detection algorithm to determine the collision behavior between the particle and the packaging model surface in real time. After the collision, the reflection direction is calculated based on the surface normal direction and the particle incident angle based on the law of conservation of momentum:

[0063] v 反射 =v 入射 -2(v 入射 ·n)n

[0064] Among them, v 反射 is the velocity vector of the particle after collision, which is determined by the incident velocity and the surface normal direction, v 入射 is the velocity vector of the particle before collision, representing the direction and speed of the particle in three-dimensional space, and n is the surface normal vector at the collision point. Energy attenuation coefficients (such as friction coefficient and elastic coefficient) are also introduced to adjust the velocity and distribution of the particles after sputtering.

[0065] Step 203: Dynamically adjust the color and transparency of the particles according to the kinetic energy attenuation during the particle movement (caused by collision and air resistance):

[0066] S2031. Initially, high-kinetic energy particles adopt bright tones (such as white or light colors), and gradually transition to dark tones (such as dark gray or black) as the energy decays to simulate the energy dissipation effect.

[0067] S2032. The lower the energy, the higher the particle transparency (such as linear interpolation based on the energy value), simulating the particle dissipation or deposition process.

[0068] When using, combine the contents in steps 201 to 203:

[0069] By combining physical simulation with numerical calculation, the texture generation process is endowed with realistic dynamic characteristics, allowing the final texture to present natural light and shadow transitions, energy attenuation traces, and collision details on complex packaging surfaces, significantly enhancing the visual immersion of the three-dimensional packaging model and the credibility of industrial simulation.

[0070] Step 3: Use ray tracing and real-time rendering technology to globally optimize the 3D model. During the rendering process, simulate complex light and shadow, transparency, reflection, and refraction to enhance the realism and detail of the packaging design.

[0071] The step three includes the following steps:

[0072] Step 301: Model the light reflection characteristics of different materials based on the BRDF (bidirectional reflectance distribution function) model; first calculate the reflection characteristics of a beam of light from the incident direction w i When irradiating a point on the surface, from the outgoing direction w o The ratio of the reflected light brightness to the incident light intensity is then used to model the light reflection characteristics of different materials based on the calculated ratio. The expression is:

[0073]

[0074] Among them L o is the initial brightness, E i is the incident irradiance.

[0075] Step 302: Distinguish primary light, diffuse light, and secondary reflection light through light layering technology, and perform independent rendering optimization on each layer of light.

[0076] Step 303: Combining a transparency processing algorithm based on volume rendering to improve the visual performance of transparent materials.

[0077] When using, combine the contents in steps 301 to 303:

[0078] By using ray tracing and real-time rendering technology to globally optimize the 3D model and simulating complex light and shadow, transparency, reflection and refraction during the rendering process, this method enhances the realism and detail of the packaging design.

[0079] Step 4: Through multimodal fusion technology, combined with text collaborative prompt learning algorithm, the text description entered by the user is parsed into a high-dimensional semantic vector and an adaptive design template is generated; further combined with image input, the generated results are optimized through cross-modal matching.

[0080] The step 4 includes the following steps:

[0081] Step 401: Process the text description and image content input by the user using a multi-layer neural network, automatically screen image features related to the design goal using cross-modal matching technology, and combine it with a gradient descent-based optimization algorithm to generate a design template that meets diversity and aesthetic requirements. Specifically, the process includes:

[0082] refer to Figure 3 , we can see that the multimodal technology process is as follows:

[0083] S4011. Use a deep learning-based natural language processing model to perform multi-granular semantic analysis on the text description entered by the user to extract keywords, style preferences, and design constraints; use a pre-trained multi-layer Transformer network to encode the parsed semantic information into a high-dimensional semantic vector, and construct a quantifiable semantic representation space that includes color tendencies, layout features, and element associations.

[0084] S4012. Establish a dual-channel deep neural network architecture, in which the text processing branch uses the self-attention mechanism to dynamically assign weights to semantic vectors, and the image processing branch extracts the texture features, composition patterns, and style elements of the input image through a convolutional neural network; introduce a cross-modal attention gating module in the feature fusion layer to realize dynamic screening of image features under the guidance of text semantics, and construct a cross-modal joint feature space with semantic consistency.

[0085] S4013. Construct a learnable similarity measurement matrix in the joint feature space, and calculate the matching degree between the text semantic vector and the image feature vector through a contrastive learning algorithm; adopt content-aware feature recombination technology to perform adaptive weighted fusion of image features that meet the preset similarity threshold to generate a parameterized template representation of the preliminary design scheme.

[0086] S4014. Establish a dual-objective optimization framework that includes an aesthetic evaluation model and a diversity discriminator. The aesthetic evaluation model quantifies the visual coordination of design solutions through deep metric learning, and the diversity discriminator adopts an adversarial training mechanism to ensure the innovativeness of the generated results. Design an iterative optimization algorithm based on improved gradient descent to simultaneously optimize the style consistency loss, aesthetic score loss, and diversity adversarial loss in the parameter space, and generate the optimal design template that meets user needs through multiple rounds of parameter adjustment.

[0087] When using, combine the content in step 401:

[0088] Through multimodal fusion technology, combined with text collaborative prompt learning algorithm, the text description entered by the user is parsed into a high-dimensional semantic vector, and an adaptive design template is generated, breaking through the semantic limitations of traditional single-modal analysis and achieving two-way deep alignment of text descriptions and image features; further combined with image input, the generation results are optimized through cross-modal matching, and the generated solution has both logical rationality and creative divergence, avoiding the homogeneity defects of traditional template generation technology.

[0089] Step 5: Integrate the diffusion model and NeRF (Neural Radiance Field) technology to generate a highly detailed 3D packaging model. NeRF is used to model the real light field environment, and the diffusion model is used to improve the diversity and clarity of texture generation.

[0090] The step five includes the following steps:

[0091] refer to Figure 4 , we can know the specific process of integrating diffusion model and NeRF (neural radiation field) technology:

[0092] Step 501: Using the staged generation strategy of the diffusion model, a rough texture layout is first generated, and then the texture details are gradually improved through multi-step denoising, specifically including:

[0093] S5011. Construct a generative architecture based on the denoising diffusion probability model (DDPM). In the initial stage, the design template is mapped to a low-resolution latent space through latent space dimensionality reduction technology; using a trainable noise prediction network, a rough texture layout map containing global composition information, basic color block distribution and key element positioning is generated according to user-specified style constraints.

[0094] S5012. A staged denoising strategy is adopted to gradually improve the resolution in the reverse process of the diffusion model. First, the noise of the low-dimensional potential representation is estimated through the U-Net network structure, and the mid-frequency texture features (such as material surface characteristics and light and shadow transition trends) are restored in combination with the cross-scale feature fusion module; then the spatial adaptive attention mechanism is introduced to inject the cross-modal semantic vector output by step 4 into the high-level feature map to guide the directional generation of high-frequency details (such as decorative patterns, edge sharpness, and micro-texture); further, a dynamically attenuated noise scheduling coefficient is set, and the continuous evolution of texture details from blur to clarity is achieved through multi-step Markov chain updates.

[0095] S5013. At each time step of the diffusion model, the gated cross-attention layer is used to calculate the correlation between the key attributes of the text description (such as "retro metal texture") and the current generated area. Then, the residual connection technology is used to integrate the semantic conditional information into the noise prediction network, and local texture enhancement is implemented for specific areas (such as buttons and icons) to ensure the consistency of detail generation with design intent.

[0096] S5014: Monitor the frequency domain energy distribution and structural similarity index of the generated image in real time. When the high-frequency detail entropy value reaches the preset aesthetic evaluation threshold or the improvement amplitude of continuous iterations is less than the critical value, the denoising process is automatically terminated, and then the final texture design scheme with both global coordination and local refinement is output.

[0097] Step 502: Combine the Mip-NeRF technology and use multi-resolution sampling to reconstruct the light field to ensure the texture consistency of the model under multi-angle observation, which specifically includes:

[0098] S5021. Reconstruct the geometric outline and basic radiation properties at low resolution (32×32), then fuse the high-frequency texture generated by diffusion at a high-resolution (128×128) layer, and combine it with differentiable volume rendering to generate a multi-view anti-aliased image.

[0099] S5022. Introduce a hierarchical feature decoupling mechanism to separate the density field and the texture radiation field into models and couple them through the physical rendering equation. During the ray travel process, the sampling interval is dynamically adjusted according to the local curvature, and super-sampling is performed on the high curvature edge area to maintain texture sharpness.

[0100] S5023. Establish a 3D-2D joint optimization loop. By comparing the structural similarity between the virtual view image and the diffusion generation result, reversely optimize the NeRF geometric field to eliminate texture breaks caused by self-occlusion and ensure visual consistency under multi-angle observation.

[0101] Step 503: Optimize the accuracy of texture generation and light field modeling through a joint loss function to improve the clarity and realism of the three-dimensional packaging model.

[0102] When using, combine the contents in steps 501 to 503:

[0103] The diffusion model's multi-stage denoising generation mechanism gradually enhances the clarity of surface microtextures (such as material particles and printed patterns). NeRF's multi-resolution anti-aliasing light field modeling ensures texture consistency across different viewing angles for 3D models. The diffusion model's high-frequency detail generation capabilities are deeply coupled with NeRF's physically-based rendering equations, resolving the inherent conflict between texture projection distortion (such as UV stretching) and view-dependent artifacts (such as flicker noise) in traditional methods. This ensures that 3D packaging models display sub-pixel detail fidelity at all viewing angles.

[0104] Step 6: Support real-time adjustment of generated results. Users can dynamically modify the model's geometry, texture style, lighting effects, and texture density through the interactive interface, and optimize the design through a real-time feedback mechanism. The interactive interface provides users with flexible operation tools by combining multiple methods such as clicking, dragging, and numerical input. It supports real-time preview of the specific impact of different adjustment parameters on the design effect, and saves the adjustment path through the intelligent history record function for subsequent review or improvement.

[0105] In order to verify the feasibility and superiority of the technical solution of the present invention, experimental analysis was carried out based on the following examples:

[0106] Example: Modeling efficiency and rendering quality comparison experiment

[0107] Experimental purpose: To verify the significant advantages of the three-dimensional packaging modeling method proposed in this invention over traditional methods in terms of modeling efficiency and rendering quality.

[0108] Test case:

[0109] Test model: irregular glass bottle (high curvature surface, translucent material).

[0110] Comparison method: Traditional parametric modeling (manual modeling and texturing in Maya);

[0111] The method of the present invention (full process automatic generation).

[0112] Experimental conditions: Hardware configuration: CPU Intel i7-12700K, GPU NVIDIA RTX 3080;

[0113] Software environment: Unity 2022.3.1f1, PyTorch 1.12.1.

[0114] Experimental steps:

[0115] 1. Modeling process:

[0116] Traditional method: Designers manually define the geometric topology, add texture maps layer by layer, and record the total time taken;

[0117] The method of the present invention includes: generating the geometry and texture framework through the Gaussian sputtering algorithm in step 1, optimizing the particle distribution through physical simulation in step 2, generating the final model through the diffusion model and NeRF in step 5, and recording the time consumption of the whole process.

[0118] 2. Rendering quality assessment:

[0119] Subjective scoring: Five professional designers were invited to score the models generated by the two methods (1-10 points), evaluating the realism of light and shadow, texture details, and material consistency.

[0120] Objective indicator: Measures the refractive index deviation of glass material (compared to the refractive index of real glass 1.52).

[0121] Experimental results:

[0122] Table 1 Comparison of modeling efficiency

[0123]

[0124] Conclusion: As shown in Table 1, the modeling efficiency of the present invention is improved by 13 times, and the automated process significantly shortens the design cycle.

[0125] Table 2 Rendering quality scores

[0126]

[0127] Designer feedback: As shown in Table 2, it can be seen that the light and shadow transition of the model of the present invention is natural, the glass refraction effect is realistic, and the texture details (such as the bottle mouth relief) are clear.

[0128] Table 3 Physical consistency verification

[0129]

[0130] Conclusion: As shown in Table 3, through the physical simulation in step 2 and the NeRF technology in step 5, the material property deviation is controlled within 3%, which meets the industrial design accuracy requirements.

[0131] Experimental Conclusion

[0132] Improved efficiency: This invention reduces modeling time from 145 minutes to 11 minutes through the Gaussian sputtering algorithm and automated process;

[0133] Quality optimization: The rendering quality score reached 9.2 points, and the refractive index deviation was only 2.1%, which is significantly better than traditional methods;

[0134] Technical advantages: The ray tracing in step 3 is combined with the NeRF technology in step 5 to achieve highly realistic light field and material performance.

[0135] Note: The experimental data is based on a standardized testing environment to ensure repeatability. The technical solution of this invention has been verified in multiple complex packaging design scenarios and has demonstrated significant technical advantages.

[0136] It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope limited by the appended claims of the application.

Claims

1. A three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation, characterized in that: The following steps are involved: Step 1: Generate the geometric structure and surface texture of the three-dimensional packaging model using Gaussian distribution to obtain particle distribution; Step 2: Based on the particle distribution and combined with the physical simulation technology based on the particle system, a particle attribute data set containing physical consistency parameters is generated; Step 3: Based on the particle attribute dataset, ray tracing and real-time rendering technology are used to perform global optimization of the 3D model; Step 4: Combine the rendered 3D model features with the multimodal data input by the user and use the text collaborative prompt learning algorithm to generate an adapted design template; Step 5: Integrate the diffusion model and Neural Radiation Field (NeRF) technology to generate a detailed 3D packaging model based on the physical consistency parameters and the modeling results of Step 3, combined with the design template.

2. The three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation according to claim 1, characterized in that: The step 1 is specifically implemented as follows: using a Gaussian distribution model to generate the geometric structure and surface texture of a three-dimensional packaging model, controlling the distribution density and range of particles in space by using the mean, standard deviation, and sputtering angle; and adjusting the ions by combining dynamic parameter adjustment technology; The particles are discrete three-dimensional basic units generated based on the Gaussian sputtering algorithm, and carry position, color, and energy attributes.

3. The three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation according to claim 2, characterized in that: The specific implementation process of step 2 is as follows: First, environmental parameters are introduced to simulate the acceleration changes of particles in three-dimensional space; Secondly, the collision detection algorithm is used to simulate the collision behavior between particles and the model surface and calculate the particle distribution path after sputtering; Finally, according to the energy attenuation law of particle motion, the particle color and transparency are adjusted to generate a realistic texture effect.

4. The three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation according to claim 3, characterized in that: The specific implementation process of step 3 is as follows: Step 3.1: Model the light reflection characteristics of different materials based on the bidirectional reflectance distribution function (BRDF) and establish a mathematical relationship between the material surface reflectivity, roughness, and incident / exit angles. Step 3.2: Based on the BRDF modeling results, use the light layering technology to divide the light into primary light, diffuse light, and secondary reflection light, and design a rendering strategy for each light: Step 3.3: Combining the light reflection characteristics of different materials with the light layering results, a transparency processing algorithm based on volume rendering is used to simulate the light transmission and internal scattering effects of transparent materials through the ray absorption-scattering model, and the light path integral is optimized according to the BRDF refractive index parameters, ultimately achieving a visual fusion of multi-layer refraction and surface reflection of transparent materials.

5. The three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation according to claim 4, characterized in that: The rendering strategy for each light is as follows: Main ray: directly calculates specular reflection and Fresnel effect in combination with BRDF parameters; Diffuse reflection: Optimize sampling density based on material roughness parameters to improve noise suppression efficiency in diffuse reflection areas. Secondary reflected rays: For reflected rays, computing resources are dynamically allocated based on the BRDF specular weight.

6. The three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation according to claim 5, characterized in that: The specific implementation of step 4 is as follows: combining the rendered three-dimensional model features with the multimodal data input by the user, combining with the text collaborative prompt learning algorithm, parsing the text description input by the user into a semantic vector, and using cross-modal matching technology to align the rendering results with user needs, automatically screening image features related to the design goals, and combining with the gradient descent-based optimization algorithm to generate an adapted design template.

7. The three-dimensional packaging modeling method based on three-dimensional Gaussian sputtering representation according to claim 6, characterized in that: The specific implementation process of step 5 is as follows: Using the staged generation strategy of the diffusion model, a rough texture layout is first generated, and then the texture details are gradually improved through multi-step denoising to obtain a refined texture layout; For the refined texture layout, the light field is reconstructed using multi-resolution sampling combined with the multi-scale representation of the neural radiance field Mip-NeRF technology; The accuracy of texture generation and light field reconstruction is optimized through a joint loss function.

Images