Physical rendering image generation method based on material estimation and illumination estimation
By embedding material parameters into three-dimensional Gaussian atoms and constructing a lighting estimation network, the problem of efficient rendering but lack of interpretable materials in explicit reconstruction methods is solved. This achieves efficient material extraction and stable training, generating interpretable physically rendered images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-03-13
AI Technical Summary
Existing 3D reconstruction methods are efficient in rendering explicit representations but lack interpretable materials, while implicit representations are deficient in terms of material interpretability and training stability, making it difficult to achieve both simultaneously.
By embedding primary color, metallicity, and roughness material parameters into three-dimensional Gaussian atoms, a material estimation and lighting estimation network is constructed. Using differentiable Gaussian rasterization and the Disney Physical Shading Model, physically rendered images are generated directly in the pixel domain, avoiding the mixing of lighting information into the material.
It enables the acquisition of object material information based on fast explicit rendering, improving extraction efficiency, and does not rely on external priors, thus ensuring the interpretability of material parameters and training stability.
Smart Images
Figure CN121661227A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to 3D reconstruction technology and deep learning technology, and in particular to a physically based rendering image generation method based on material estimation and lighting estimation. Background Technology
[0002] Multiview Figure 3 Research in 2D reconstruction can be broadly categorized into two types: one is implicit neural representation centered on volume rendering (such as the NeRF series), which maps continuous functions (color / density or SDF, etc.) in space to 2D images through differentiable volume integrals; the other is explicit representation centered on geometric primitives (such as 3D Gaussian Splatting, or 3DGS for short), which uses differentiable Gaussian ellipsoids for weighted synthesis in screen space to obtain images from new perspectives. The former has the advantage of expressing continuous fields and supporting end-to-end optimization, but its rendering efficiency is relatively low, and it couples all material, lighting, and geometric information into the implicit network, resulting in weak material interpretability; the latter has the advantage of high rendering efficiency and direct manipulation of primitive properties, but mainstream work focuses on improving reconstruction quality and view synthesis accuracy, lacking systematic modeling and supervision of recoverable material properties (such as primary color / metallicity / roughness).
[0003] Existing material estimation work typically relies on: (1) additional priors (environment mapping, known lighting) or complex decomposition assumptions; (2) "end-to-end fitting" of materials and lighting into the same implicit network, which can achieve good reconstruction on the image, but often lacks stable and interpretable supervision signals in the material dimension; (3) relying on explicit meshing and then mapping, which has a long process and a long gradient chain, making it difficult to reconcile with the fast differentiable rendering advantages of 3DGS.
[0004] Overall, existing technologies struggle to simultaneously achieve "differentiable rendering, material interpretability, and training stability": either training converges to near-pure appearance regression at the expense of physical meaning, or complex processes, additional annotations, or cumbersome geometric intermediates are introduced to ensure interpretability, making it difficult to efficiently unify them into single-branch training in 3DGS. Summary of the Invention
[0005] To address the issue of "efficient rendering but lack of interpretable materials" in explicit reconstruction methods such as 3DGS, this invention proposes a physically based rendering image generation method based on material estimation and lighting estimation, specifically including the following steps:
[0006] A set of three-dimensional Gaussian atoms is initialized based on multi-view images and camera intrinsic and extrinsic parameters. Three types of material parameters, namely primary color, metallicity and roughness, are added to each Gaussian atom.
[0007] Under any training perspective, the color synthesis of the current frame is completed through a differentiable Gaussian rasterization process. At the same time, the renderer synchronously returns the pixel-Gaussian mapping matrix of each pixel, that is, the Gaussian domain information is projected into the pixel domain.
[0008] A material estimation network from material to spherical harmonic coefficients is constructed, taking Gaussian ellipsoid material, viewpoint, and position as inputs, and embedding material properties into spherical harmonic coefficients;
[0009] An ambient light estimation network is constructed to obtain current ambient light information. The SH coefficient of the Gaussian ellipsoid, the position, and the viewing angle are used as inputs to obtain the main incoming direction and light intensity under the viewing angle.
[0010] Based on the obtained material and lighting estimates, the Disney Physical Shading Model is used to directly use the projected primary color as the diffuse primary color in the pixel domain, and combine metallicity, roughness, normal and incident light to generate a physically rendered image.
[0011] Compared to 3D reconstruction methods in recent years, the method proposed in this invention can simultaneously acquire the material information of objects while performing fast explicit rendering, without relying on external priors. Compared to implicit methods that focus on material extraction, the extraction efficiency is significantly improved, while compared to explicit material extraction methods, it can still achieve independence from external prior input. Attached Figure Description
[0012] Figure 1 This is a schematic diagram of a physically based rendering image generation method based on material estimation and lighting estimation according to the present invention.
[0013] Figure 2 This is a schematic diagram of the material estimation network in this invention;
[0014] Figure 3 This is a schematic diagram of the ambient light estimation network in this invention;
[0015] Figure 4 This is a visual representation of the material extraction process from the TNT dataset in this invention.
[0016] Figure 5 This is a comparison result of the present invention on the DTU dataset scan97 data. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] This invention proposes a physically based rendering image generation method based on material estimation and lighting estimation, specifically including the following steps:
[0019] A set of three-dimensional Gaussian atoms is initialized based on multi-view images and camera intrinsic and extrinsic parameters. Three types of material parameters, namely primary color, metallicity and roughness, are added to each Gaussian atom.
[0020] Under any training perspective, the color synthesis of the current frame is completed through a differentiable Gaussian rasterization process. At the same time, the renderer synchronously returns the pixel-Gaussian mapping matrix of each pixel, that is, the Gaussian domain information is projected into the pixel domain.
[0021] A material estimation network from material to spherical harmonic coefficients is constructed, taking Gaussian ellipsoid material, viewpoint, and position as inputs, and embedding material properties into spherical harmonic coefficients;
[0022] An ambient light estimation network is constructed to obtain current ambient light information. The SH coefficient of the Gaussian ellipsoid, the position, and the viewing angle are used as inputs to obtain the main incoming direction and light intensity under the viewing angle.
[0023] Based on the obtained material and lighting estimates, the Disney Physical Shading Model is used to directly use the projected primary color as the diffuse primary color in the pixel domain, and combine metallicity, roughness, normal and incident light to generate a physically rendered image.
[0024] This invention aims to address the pain point of explicit reconstruction methods such as 3DGS, which are "efficient in rendering but lack interpretable materials." This invention seeks to directly generate material parameter maps within a single-branch training framework without relying on volumetric rendering integrals, mesh intermediates, or external lighting calibration, ensuring consistency with pixel rendering results and subjecting them to the stability constraints of the physically based rendering equations. This invention chooses to embed pure material information into a low-frequency interpretable spherical harmonic space (SH), and then explicitly regresses the lighting direction and intensity from this SH expression, avoiding the miswriting of lighting into materials. Simultaneously, it utilizes the pixel-level visibility inherent in 3DGS rasterization to aggregate Gaussian domain materials into a material map in screen space, ensuring that the training objective and the rendering process are of the same origin.
[0025] This invention explicitly embeds a learnable material vector (primary color / metallicity / roughness) into each 3D Gaussian. During rendering, the pixel Gaussian mapping matrix obtained by the rasterizer is used to perform differentiable aggregation of materials in screen space, directly obtaining a material map aligned pixel-by-pixel with the image. Based on this, to avoid the ambiguity of writing lighting into the material training, a lightweight material estimation network is first introduced to embed the pure material into a low-order spherical harmonic space and establish a low-frequency consistency constraint with the spherical harmonic coefficients of the rendering branch. Then, the lighting estimation network explicitly regresses the principal input direction and intensity from this low-frequency expression, extracting the external lighting as controllable parameters. Finally, the material map and estimated lighting are substituted into a simplified Disney physical rendering model to synthesize the predicted image in the pixel domain.
[0026] This invention addresses explicit estimation of material parameters and physically consistent rendering under multi-view conditions. It employs a three-dimensional Gaussian field as the basic representation and directly embeds a material triple (primary color / metallicity / roughness) into each Gaussian atom to achieve explicit controllability and learnability of appearance factors in space. The specific process of this invention is as follows: Figure 1 Specifically, it includes:
[0027] First, a set of three-dimensional Gaussian atoms is initialized based on multi-view images and camera internal and external parameters. For example, 3D Gaussian splashing technology is used to generate three-dimensional Gaussian atoms based on the input image. In addition to classic parameters such as center position μ, scale, rotation and opacity α, the generated three-dimensional Gaussian atoms are supplemented by three material parameters for each Gaussian atom: base color A (BaseColor / Albedo), metallicity M (Metallic), and roughness R (Roughness).
[0028] Subsequently, under any training viewpoint, the color synthesis of the current frame is completed through a differentiable Gaussian rasterization process. At the same time, the renderer synchronously returns the pixel-Gaussian mapping matrix of each pixel, thereby efficiently projecting Gaussian domain information into the pixel domain in a stable manner without relying on volume rendering integrals or mesh reconstruction.
[0029] Regarding material decoupling, a material estimation network Net is first constructed to estimate the material to the spherical harmonic (SH) coefficients. SH2L The material (including primary color A, metallicity M, roughness R), viewpoint (including CameraDir, which is a vector used to define the camera orientation in 3D graphics rendering, usually represented as a unit vector pointing from the camera position to the viewing direction) and position (the center position μ of the 3D Gaussian atom and the normal vector n of the 3D Gaussian atom) are used as inputs. The material properties and other parameters are embedded into the spherical harmonic coefficients. That is, the material estimation network estimates the spherical harmonic (SH) coefficients based on the input data.
[0030] Then construct the ambient light estimation network Net. M2SHThe network obtains the current ambient lighting information, and uses the SH coefficient of the Gaussian ellipsoid, position and viewing angle as input to obtain the main incoming direction and light intensity under the viewpoint, so that the rendering equation can still be physically interpreted under the condition of no external light source information input.
[0031] Based on the above material and lighting estimation, a simplified Disney physical shading model is adopted. In the pixel domain, the primary color obtained by projection is directly used as the diffuse primary color. Combined with metallicity, roughness, normal and incident light, a physically rendered image is generated.
[0032] During the training phase, the error between the rendered image and the real image is minimized, and smoothing and physical extent priors are applied to the material map and normal map, so that the material parameters, lighting parameters and Gaussian representation converge in a coordinated manner throughout the end-to-end process, thereby further improving the material resolution and cross-view consistency.
[0033] Specifically, this embodiment bundles material and visibility into the same carrier, that is, the material parameters are directly embedded into the three-dimensional Gaussian ellipsoid itself. The direct benefit of this is that visibility and transparency information from differentiable rasterization paths will participate in the material's projection, supervision, and backpropagation with the same entity index and the same set of weights. Therefore, in this invention, the first... The number of Gaussian atoms is:
[0034]
[0035] in, It is the first The three-dimensional center position of a three-dimensional Gaussian atom. For the real number space, The superscript indicates the dimension of the real number space; It is the first The three-dimensional anisotropic covariance of a three-dimensional Gaussian atom is given by the first... The scale vector of a three-dimensional Gaussian atom and the Rotation matrix of three-dimensional Gaussian atoms Parameterization; The first one that is constrained by the Sigmoid constraint The opacity of a three-dimensional Gaussian atom; It is the first RGB primary color of a three-dimensional Gaussian atom (also known as Albedo / BaseColor); For the first The metallicity of a three-dimensional Gaussian atom; For the first Material roughness of a three-dimensional Gaussian atom (to avoid interference with geometric rotation) (Note: The roughness is indicated by the superscript "mat" for distinction). These parameters determine the elliptical projection shape and composition order of the Gaussian ellipsoid on the screen, and also participate in the subsequent physical rendering equations as learnable material parameters.
[0036] The above method can embed interpretable material parameters into a unified Gaussian representation. However, pure material properties themselves do not carry ambient lighting information. Directly using material parameters to fit pixel colors can easily lead to incorrect attribution of low-frequency signal differences to material or lighting information, resulting in abnormal training performance. Therefore, this invention designs a material estimation network that estimates material parameters, i.e., the three-dimensional center positions of three-dimensional Gaussian atoms. The material parameters are: RGB primary color A, metallicity M, material roughness R, and the normal direction n of the current 3D Gaussian atom. The spherical harmonic coefficients SH are estimated. During training, the estimated spherical harmonic coefficients are compared with the spherical harmonic coefficients returned by the raster renderer, so that the material parameters are indirectly involved in the training.
[0037] like Figure 2 In this invention, the material estimation network comprises eight cascaded fully connected linear layers, each with a dimension of 128. The last fully connected linear layer outputs the estimated spherical harmonic coefficients. Two skip connections and one injection are also included in the eight cascaded fully connected linear layers. The first skip connection concatenates the input data of the material estimation network with the output of the second fully connected linear layer as the input of the third fully connected linear layer. The second skip connection concatenates the output of the fourth and sixth fully connected linear layers as the input of the fifth fully connected linear layer. Injection refers to concatenating the camera's viewing direction with the output of the third fully connected linear layer as the input of the fourth fully connected linear layer.
[0038] For each Gaussian ellipsoid, the spherical harmonic coefficients estimated by this invention are as follows:
[0039]
[0040] in, These are the estimated spherical harmonic coefficients. For material estimation network; It represents the number of color channels, usually 3, and B is the number of spherical harmonic coefficients expanded. This is the material vector of the current Gaussian ellipsoid; It is the viewpoint direction, i.e., CameraDir, which is used as a parameter for injecting material estimation networks; This is the current three-dimensional position corresponding to Gaussian. This is the direction of the normal to the current Gaussian (i.e., the direction of the minimum scaling factor). The spherical harmonic coefficients obtained above... With raster renderer return coefficients The contrastive loss is used for model training, primarily for training the material estimation network.
[0041]
[0042] in, Spherical harmonic coefficients With raster renderer return coefficients The contrast loss between the spherical harmonic coefficients output by the material estimation network and the L1 norm between the coefficients returned by the raster renderer; This represents L1 norm operations.
[0043] In physically based rendering, incident light (direction and intensity) determines the coupling method between diffuse and specular reflection in terms of normals, viewpoints, and material parameters. If lighting is not explicitly estimated but implicitly embedded in materials or other parameters, lighting information will be missing during the final physically based rendering, forcing material information to bear the burden of color changes interpreted by lighting information. Since spherical harmonics are a compact encoding of the low-frequency distribution of the environment on a sphere, under diffuse and specular reflection conditions, the principal direction and energy (i.e., direction and intensity) information of incident light will be concentrated in low-order and some mid-order spherical harmonics. Therefore, this invention designs a lightweight network for estimating lighting information using spherical harmonic functions, namely, an ambient light estimation network, such as... Figure 3 The network consists of four cascaded linear fully connected layers with dimensions of 256, 128, 64, and 64, respectively. Two residual connections are set up. The first residual connection adds the data input to the ambient light estimation network to the output of the first-level linear fully connected layer as the input to the second-level linear fully connected layer. The second residual connection adds the output of the second-level linear fully connected layer to the output of the third-level linear fully connected layer as the input to the fourth-level linear fully connected layer. The output of the fourth layer is mapped to 3D illumination direction and 1D illumination intensity through two parallel linear fully connected layers.
[0044] Ambient light estimation network for spherical harmonic coefficients SH and Gaussian position and the direction of observation As input, we obtain the direction and intensity of the light. , is represented as:
[0045]
[0046] in, The direction of the light; Light intensity; A network for estimating ambient light.
[0047] In Gaussian differentiable rasterization, the color of each pixel p is obtained by superimposing a ray along the viewing direction with a set of ordered Gaussians. For the Gaussians hit by the ray, the first... The contribution of each three-dimensional Gaussian atom to pixel p is:
[0048]
[0049] in, It is the first The opacity of a three-dimensional Gaussian atom. Is it to reach the The cumulative transmittance before a three-dimensional Gaussian atom.
[0050] When the renderer performs color overlay, this invention simultaneously records the ID matrix of the top K Gaussian values that contribute the most to each pixel under the current viewpoint. and the corresponding contribution matrix This yields pixel-level mapping matrices, all of which are 3D matrices H*W*K, where H and W represent the height and width of the corresponding pixel image, and K represents the top K Gaussian weights in terms of total contribution (i.e., weights), sorted from largest to smallest. Based on the mapping matrix, the corresponding material image (single pixel p) at the current viewpoint can be obtained:
[0051]
[0052] in, The weighted cumulative sum of the primary colors at pixel p is calculated by weighting the contribution values of the top K primary colors with the highest contribution values from the three-dimensional Gaussian atoms. This represents the corresponding i-th element in the current mapping matrix (i.e., the matrix composed of the top K three-dimensional Gaussian atoms that contribute to the current pixel). A three-dimensional Gaussian atom; The weighted cumulative metallicity at pixel p is calculated by weighting the metallicity of the top K highest contribution values by the contribution values of the three-dimensional Gaussian atoms. The weighted cumulative roughness at pixel p is calculated by weighting the roughness cumulative values of the top K roughness values with the highest contribution values from the three-dimensional Gaussian atoms.
[0053] After material extraction, physically based rendering can be performed by combining the obtained ambient lighting information. This invention uses a simplified Disney model to perform physically based rendering in the pixel domain, with the light emission intensity based on the following rendering equation:
[0054]
[0055] Where x represents the ray in the direction of observation (outgoing direction) The coordinates of the point hit can be understood as pixels in the rendered image, and n represents the normal vector at that point. Represents the bidirectional reflection distribution function (BRDF). Represents the direction of incident light. Representative with The direction is the center of the hemispherical domain; It comes from the direction The incident light intensity. In this invention, the ambient light and its intensity are estimated using a light estimation network, thus transforming the formula into a non-integral form:
[0056]
[0057] According to the Disney model, the BRDF function can be decomposed into two terms: diffuse reflection and specular reflection, namely:
[0058]
[0059] The diffuse reflection term is represented as , This is the weighted cumulative sum of metallicity at pixel x. The weighted cumulative sum of the primary colors at pixel x; mirror term. Then the Cook-Torrance form is adopted, based on the normal distribution function. Geometric functions and Fresnel items Composition, namely:
[0060]
[0061] in, This represents the half-length vector between the observation direction and the incident direction, that is, the midpoint of the angle between the two directions. Represents absolute value operation; Related to the previous text Equivalent, both representing the material properties of the current pixel; the normal distribution function D represents the probability density that the micro-surface orientation is consistent with the direction of the half-length vector; the Fresnel term F describes the change in reflectivity when entering the material at different incident angles; the geometric function G describes the self-occlusion and shadow occlusion between micro-surfaces, and the corresponding formulas are as follows:
[0062]
[0063]
[0064]
[0065] in, To and and The relevant intermediate parameters are represented as follows:
[0066]
[0067] For a with and The relevant intermediate function is represented as follows:
[0068]
[0069] in or .
[0070] Finally, this invention performs supervised comparison between the obtained rendered image and the real image to ensure the effectiveness of the training results:
[0071]
[0072] in, The contrast loss between the rendered image and the real image; , As a weighting factor, in this embodiment The value is 0.3. The value is 0.7; With the direction of incident light The value of the generated physically rendered image at pixel position x, where n is the normal vector at that pixel position; This represents the value of the actual image at pixel position x. This represents a similarity calculation used to measure the structural similarity between a rendered image and a real image. This metric comprehensively reflects whether the brightness, contrast, and local structure of the images are consistent.
[0073] To verify the effectiveness of the method proposed in this invention, this embodiment evaluates the performance of the method on the TNT (Tanks and Temples) dataset. The test results are as follows: Figure 4 We selected data (scan_97) from the DTU dataset for comparison with other methods, and the comparison results are as follows: Figure 5 The model in this invention uses the AdamW optimizer (…). The learning rate is managed by a configuration file and by default uses cosine annealing to smoothly decrease from the initial learning rate to the minimum learning rate. Training is performed with multi-view images as input.
[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A physically based rendering image generation method based on material estimation and lighting estimation, characterized in that, Specifically, the following steps are included: A set of three-dimensional Gaussian atoms is initialized based on multi-view images and camera intrinsic and extrinsic parameters. Three types of material parameters, namely primary color, metallicity and roughness, are added to each Gaussian atom. Under any training perspective, the color synthesis of the current frame is completed through a differentiable Gaussian rasterization process. At the same time, the renderer synchronously returns the pixel-Gaussian mapping matrix of each pixel, that is, the Gaussian domain information is projected into the pixel domain. A material estimation network from material to spherical harmonic coefficients is constructed, taking Gaussian ellipsoid material, viewpoint, and position as inputs, and embedding material properties into spherical harmonic coefficients; An ambient light estimation network is constructed to obtain current ambient light information. The SH coefficient of the Gaussian ellipsoid, the position, and the viewing angle are used as inputs to obtain the main incoming direction and light intensity under the viewing angle. Based on the obtained material and lighting estimates, the Disney Physical Shading Model is used to directly use the projected primary color as the diffuse primary color in the pixel domain, and combine metallicity, roughness, normal and incident light to generate a physically rendered image.
2. The physically based rendering image generation method according to claim 1, characterized in that, A material estimation network from material to spherical harmonic coefficients is constructed. This network consists of eight cascaded fully connected linear layers, each with a dimension of 128. The last fully connected linear layer outputs the estimated spherical harmonic coefficients. Two skip connections and one injection are also included in the eight cascaded fully connected linear layers. The first skip connection concatenates the input data of the material estimation network with the output of the second fully connected linear layer as the input of the third fully connected layer. The second skip connection adds the outputs of the fourth and sixth fully connected linear layers as the input of the fifth fully connected linear layer. The injection involves concatenating the camera's viewpoint direction with the output of the third fully connected linear layer as the input of the fourth fully connected linear layer.
3. The physically based rendering image generation method according to claim 1, characterized in that, The material estimation network is trained by minimizing the L1 norm between the spherical harmonic coefficients output by the material estimation network and the coefficients returned by the raster renderer.
4. The physically based rendering image generation method according to claim 1, characterized in that, The ambient light estimation network consists of four cascaded linear fully connected layers with dimensions of 256, 128, 64, and 64, respectively. Two residual connections are set up. The first residual connection is the concatenation of the input data of the ambient light estimation network with the output of the first-level linear fully connected layer as the input of the second-level linear fully connected layer. The second residual connection is the concatenation of the output of the second-level linear fully connected layer with the output of the third-level linear fully connected layer as the input of the fourth-level linear fully connected layer. The output of the fourth layer is mapped to 3D illumination direction and 1D illumination intensity through two parallel linear fully connected layers.
5. The physically based rendering image generation method according to claim 1, characterized in that, For the physically based rendering model consisting of a material estimation network, an ambient light estimation network, and the Disney physically based shading model, the training is performed to minimize the L1 norm between the spherical harmonic coefficients output by the material estimation network and the coefficients returned by the raster renderer, as well as the contrast between the rendered image and the real image. The contrast loss between the rendered image and the real image is expressed as: in, The contrast loss between the rendered image and the real image; , As a weighting factor; With the direction of incident light The value of the generated physically rendered image at pixel position x, where n is the normal vector at that pixel position; This represents the value of the actual image at pixel position x. This indicates similarity calculation.
6. The physically based rendering image generation method according to claim 1, characterized in that, Color synthesis for the current frame is performed through a differentiable Gaussian rasterization process. In Gaussian differentiable rasterization, the color of each pixel p is obtained by superimposing a ray along the viewing direction with a set of sorted Gaussian raster shapes, specifically including: Obtain all three-dimensional Gaussian atoms hit by the ray in the observation direction, and calculate the contribution of each three-dimensional Gaussian atom to pixel p. This contribution is the product of the opacity of the current three-dimensional Gaussian atom and the cumulative transmittance before reaching the current three-dimensional Gaussian atom. Calculate the top K three-dimensional Gaussian atoms with the highest contribution values, and weight each three-dimensional Gaussian atom for its primary color, metallicity, and roughness based on its contribution. The cumulative value of the weighted average of K three-dimensional Gaussian atomic primary colors, metallicity, and roughness is used as the primary color, metallicity, and roughness values of pixel p.
7. The physically based rendering image generation method according to claim 6, characterized in that, Using a simplified Disney physically-based shading model, the primary color obtained from projection is directly used as the diffuse primary color in the pixel domain. This is combined with metallicity, roughness, normals, and incident light to generate a physically rendered image. in, This indicates that the x-axis of a pixel in a physically rendered image is along the ray direction. The value of n is the normal vector at pixel x; It is a two-way reflection distribution function. Light intensity; To provide guidance.
8. The physically based rendering image generation method according to claim 7, characterized in that, bidirectional reflection distribution function Represented as: in, For diffuse reflection; This is the weighted cumulative sum of metallicity at pixel x; The weighted cumulative sum of the primary colors at pixel x; Mirrored item; Represents the normal distribution function; Indicates Fresnel terms; Represents geometric functions; This is a weighted cumulative representation of the material roughness at pixel x; It is the half-range vector between the observation direction and the incident direction.
9. The physically based rendering image generation method according to claim 8, characterized in that, The normal distribution function is used to represent the probability density of the micro-surface orientation being consistent with the direction of the half-path vector, and is expressed as: The Fresnel term describes the change in reflectivity when the material enters at different incident angles, and is expressed as: Geometric functions are used to describe self-occlusion and shadow occlusion between micro-surfaces, and are expressed as: in, To and and The relevant intermediate parameters are represented as follows ; For a with and The relevant intermediate function is represented as follows: ,in or .
10. A physically based rendering image generation method according to claim 8 or 9, characterized in that, The half-length vector h between the observation direction and the incident direction, that is, the midpoint of the angle between the observation direction and the incident direction, is expressed as: in, This represents absolute value operations.