Method, device and electronic equipment for generating a render map

Abstract

This invention provides a rendering generation method, device, and electronic device, belonging to the field of image processing technology. It generates precise initial line drawings through parametric methods and extracts them as line drawing control diagrams, ensuring the engineering rigor of the product structure. Subsequently, in the image rendering stage, the potential representation of this line drawing is used as a strong constraint, guiding the generation process together with the design text describing the style. This avoids the randomness and unreliability of geometric structure and spatial proportions caused by pure text control, significantly improving design efficiency and rendering quality.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, apparatus and electronic device for generating rendered images. Background Technology

[0002] Industrial product design is a crucial link in the manufacturing innovation chain, and its efficiency and quality directly affect the product's market competitiveness and brand value. Traditional industrial product design processes typically include multiple stages such as conceptual sketching, detailed line drawing, and 3D modeling and rendering.

[0003] Taking rail transit vehicle design as an example, designers need to manually draw structural line drawings based on the overall parameters of the vehicle (such as body length, cross-sectional dimensions, etc.).

[0004] Traditional manual line drawing is inefficient and makes it difficult to quickly generate parametric line drawings that meet engineering specifications, which restricts the speed of design iteration. Although text-to-image generation models can generate high-quality images based on natural language descriptions, pure text control is difficult to accurately describe the geometric structure and spatial proportions of the product. The generated results often deviate from the designer's intentions and cannot obtain renderings that meet the requirements. Summary of the Invention

[0005] This invention provides a rendering generation method, device, and electronic device to address the shortcomings of traditional manual line drawing in the prior art, where the generated renderings often deviate from the designer's intentions and fail to produce satisfactory renderings. This invention achieves strong structural controllability in the industrial design rendering generation process, significantly improving design efficiency and rendering quality.

[0006] This invention provides a method for generating a rendered image, comprising the following steps.

[0007] Based on the engineering parameters of the industrial design, the initial line drawing corresponding to the industrial design is determined, and edge detection is performed on the initial line drawing to obtain the line drawing control map; The line drawing control diagram is input into a variational autoencoder for encoding to obtain a low-dimensional latent space representation. The industrial design text and the low-dimensional latent space representation are input into the image rendering model for rendering to obtain the target rendering result output by the image rendering model. The target rendering result is generated based on the industrial design text under the constraints of the low-dimensional latent space representation. The target rendering result is input into the variational autoencoder for decoding to obtain the rendering image of the industrial design.

[0008] According to a rendering image generation method provided by the present invention, the image rendering model includes an image rendering architecture and a conditional diffusion model architecture. The image rendering architecture includes a plurality of image rendering units connected in sequence, and the conditional diffusion model architecture includes an injection unit corresponding to each of the plurality of image rendering units. The image rendering units are used to generate rendering results, and the injection units are used to extract control features. During the rendering process of the image rendering model, the first image rendering unit renders the received rendering result to obtain the first rendering result; The first injection unit corresponding to the first image rendering unit performs control feature extraction on the rendering result received by the first image rendering unit to obtain a first extraction result; The first rendering result is updated based on the weight parameters of the first injection unit and the first extraction result to obtain the updated first rendering result; The updated first rendering result is sent to the second image rendering unit as the rendering result received by the second image rendering unit; The first image rendering unit and the second image rendering unit are two adjacent image rendering units.

[0009] According to a rendering method provided by the present invention, each image rendering unit includes: The self-attention layer is used to integrate information from the received rendering results to obtain the integrated result. A cross-attention layer is used to align and merge the integration result and the industrial design text to obtain a fusion result; A feedforward neural network is used to process the fusion result to obtain the rendering result output by the image rendering unit.

[0010] According to a rendering generation method provided by the present invention, the self-attention layer uses learnable gating parameters to adjust the influence ratio of the industrial design text and the low-dimensional latent space representation on the integration result; and / or The cross-attention layer uses learnable gating parameters to adjust the proportion of influence of the industrial design text and the low-dimensional latent space representation on the fusion result.

[0011] According to a rendering graph generation method provided by the present invention, the rendering graph generation method further includes: Determine the sharpening index corresponding to the updated first rendering result; Based on the sharpening index being greater than or equal to the index threshold, the rescaling factor used by the image rendering unit is reduced.

[0012] According to a rendering method provided by the present invention, the industrial design is a rail transit vehicle, and the engineering parameters include overall vehicle body parameters, front profile parameters, and component layout parameters. The determination of the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design specifically includes: Based on the aforementioned front profile parameters, the front profile curve is calculated using cubic Bézier curves. The initial line drawing corresponding to the industrial design is determined based on the vehicle front profile curve, the overall vehicle body parameters, and the component layout parameters.

[0013] According to a rendering generation method provided by the present invention, after determining the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design, and before performing edge detection on the initial line drawing to obtain the line drawing control image, the method further includes: The initial line drawing is subjected to Gaussian filtering; The Sobel operator is used to calculate the gradient magnitude and direction of the initial line drawing after Gaussian filtering; and non-maximum suppression is performed based on the gradient magnitude and the scheme. Based on a first threshold and a second threshold, strong edges and weak edges in the initial line drawing after the non-maximum suppression processing are determined, wherein the first threshold corresponds to the strong edges and the second threshold corresponds to the weak edges; The strong edge and the portion where the strong edge connects to the weak edge are retained.

[0014] According to a rendering image generation method provided by the present invention, the step of inputting the design text of the industrial design and the low-dimensional latent space representation into an image rendering model for rendering, and obtaining the rendering result output by the image rendering model, further includes: Receive a first input of a target display parameter from a plurality of display parameters, the display parameters including aspect ratio and resolution; In response to the first input, the target display parameters corresponding to the rendering result output by the image rendering model are determined.

[0015] The present invention also provides a rendering image generation apparatus, comprising the following modules: The processing module is used to determine the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design, and to perform edge detection on the initial line drawing to obtain the line drawing control map; The encoding module is used to input the line drawing control diagram into the variational autoencoder for encoding to obtain a low-dimensional latent space representation; The rendering module is used to input the design text of the industrial design and the low-dimensional latent space representation into the image rendering model for rendering, and to obtain the target rendering result output by the image rendering model. The target rendering result is generated based on the design text of the industrial design under the constraints of the low-dimensional latent space representation. The decoding module is used to input the target rendering result into the variational autoencoder for decoding to obtain the rendering image of the industrial design.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the rendering method described above.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the rendering method as described above.

[0018] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the rendering image generation method as described above.

[0019] The rendering generation method, device, and electronic equipment provided by this invention first generate a precise initial line drawing through parametric methods and extract it as a line drawing control diagram, ensuring the engineering rigor of the product structure. Subsequently, in the image rendering stage, the latent representation of this line drawing is used as a strong constraint condition, guiding the generation process together with the design text describing the style. This processing method ensures that the structural information of the output result of the image rendering model is firmly anchored by the latent representation of the line drawing when generating data, avoiding the randomness and unreliability of geometric structure and spatial proportions caused by pure text control; at the same time, the semantic guidance of the design text provides the model with rich creative space for filling in details such as materials, lighting, and colors. Ultimately, it achieves strong structural controllability in the industrial design rendering generation process, significantly improving design efficiency and rendering quality. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is one of the flowcharts illustrating the rendering image generation method provided by the present invention.

[0022] Figure 2 This is the second flowchart illustrating the rendering image generation method provided by the present invention.

[0023] Figure 3 This is a flowchart illustrating the process of determining the initial line drawing corresponding to the industrial design based on engineering parameters provided by the present invention.

[0024] Figure 4 This is the third flowchart of the rendering image generation method provided by the present invention.

[0025] Figure 5 This is the fourth flowchart of the rendering image generation method provided by the present invention.

[0026] Figure 6 This is a schematic diagram of the rendering image generation device provided by the present invention.

[0027] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0028] Figure label: 601: Processing module; 602: Encoding module; 603: Rendering module; 604: Decoding module; 710: Processor; 720: Communication interface; 730: Memory; 740: Communication bus. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] This invention provides a method for generating renderings, which can be executed by a computer device with a rendering generation application installed, such as a personal computer, workstation, server, or cloud computing platform. This method achieves controllable, efficient, and high-quality generation of industrial design renderings by using precise parametric line drawings as strong structural constraints and combining them with descriptive design text.

[0031] The following is combined Figures 1 to 7 The present invention describes a rendering method, a rendering apparatus, and an electronic device.

[0032] Figure 1 This is one of the flowcharts illustrating the rendering image generation method provided by the present invention, such as... Figure 1 As shown, the method includes the following: Step 101: Determine the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design, and perform edge detection on the initial line drawing to obtain the line drawing control map.

[0033] In some embodiments, engineering parameters refer to a series of numerical values ​​or parameter sets that can be defined in the form of quantitative data and determine the geometric shape and structural layout of an industrial product. For different industrial products, the engineering parameters can have different specific contents. For example, they can be general parameters such as length, width, and height that describe the overall size of the product, or contour parameters such as curvature, angle, and radius that define the shape of a specific part, or layout parameters such as coordinates and spacing that determine the relative positions of various components.

[0034] In some embodiments, the initial line drawing may be a vector graphic (such as SVG format) or a high-resolution bitmap image (such as PNG format), or a PNG transparent background image.

[0035] In some embodiments, the initial line drawing includes a front view and a side view, wherein the front view and the side view are displayed side by side on the same canvas to form a complete structural line drawing.

[0036] Edge detection is an image processing technique that aims to identify locations in an image where there are significant changes in brightness, color, or texture. These locations typically correspond to the edges or contours of objects. In this embodiment, edge detection refines an initial line drawing, which may contain different gray levels, line widths, or styles, into a clear, defined, and typically single-pixel-wide binary image (i.e., containing only black lines and a white background). This refined image is the line drawing control map.

[0037] In some embodiments, the Canny edge detection algorithm is used to perform edge detection to obtain the line drawing control map.

[0038] Specifically, the Canny edge detection algorithm sets the low threshold to 100 and the high threshold to 200. The extracted result is a three-channel RGB edge image with the same size as the original image, where the edge pixel value is 255 (white) and the non-edge pixel value is 0 (black).

[0039] Step 102: Input the line drawing control chart into the variational autoencoder for encoding to obtain a low-dimensional latent space representation.

[0040] The Variational Autoencoder (VAE) is a deep learning model consisting of an encoder and a decoder. In this step, the encoder part is used. The encoding process can be understood as an efficient data compression, mapping (or encoding) a high-dimensional line drawing control map containing a large amount of pixel data (e.g., a 1024x1024 pixel image) to a mathematical space with significantly reduced dimensions, namely the latent space, i.e., a low-dimensional latent space representation.

[0041] The low-dimensional latent space representation is a compact numerical vector or tensor. Although this latent representation has low dimensionality, it contains all the core structural information of the original line drawing control chart through model learning. The core purpose of this encoding step is to improve computational efficiency, because performing subsequent image rendering calculations in the low-dimensional latent space is much faster than in the original, high-dimensional pixel space, thus significantly reducing the demand for computing resources.

[0042] In some embodiments, the variational autoencoder compresses the line drawing control graph to the latent space. The encoded latent feature dimensions are [B, C, H / 8, W / 8], where B is the batch size, C is the number of channels (usually 4 or 8), H and W are 1 / 8 of the height and width of the original image, respectively. The image is then flattened into a sequence form by a pack operation for input into the image rendering model.

[0043] Step 103: Input the industrial design text and the low-dimensional latent space representation into the image rendering model for rendering, and obtain the target rendering result output by the image rendering model. The target rendering result is generated based on the industrial design text under the constraints of the low-dimensional latent space representation.

[0044] Among them, the design text of industrial design is a description entered by the user in natural language, which is used to define the style, material, color, lighting environment and other unstructured features of the rendered image, such as "silver-gray metallic body with red auspicious cloud patterns on the surface, outdoor lighting on a sunny afternoon".

[0045] In this step, the rendering result is generated based on the design text of the industrial design, under the constraints of a low-dimensional latent space representation. This means that when generating each pixel, the image rendering model must simultaneously meet two conditions: first, it must conform to the description in the design text in terms of content and style (e.g., the car body is silver-gray); second, it must strictly align with the contours defined by the latent space representation of the line art in terms of structure and shape (e.g., the car windows must appear in the positions specified in the line art). The target rendering result is the final product of this rendering process in latent space; it is a low-dimensional representation that integrates textual semantics and line art structure.

[0046] Step 104: Input the target rendering result into the variational autoencoder for decoding to obtain the industrial design rendering.

[0047] This step utilizes the decoder portion of the VAE, which functions inversely to the encoder. The decoder receives the target rendering result generated in the latent space and upsamples and reconstructs it into a high-dimensional, high-resolution pixel image. The final output image is the industrial design rendering that the user needs, conforming to both engineering structure and rich artistic detail.

[0048] In this embodiment, a precise initial line drawing is first generated parametrically and extracted as a line drawing control diagram, ensuring the engineering rigor of the product structure. Subsequently, during the image rendering stage, the latent representation of this line drawing is used as a strong constraint, guiding the generation process together with the design text describing the style. This approach ensures that the structural information of the output result of the image rendering model is firmly anchored by the latent representation of the line drawing when generating data, avoiding the randomness and unreliability of geometric structure and spatial proportions caused by pure text control. At the same time, the semantic guidance of the design text provides the model with rich creative space for filling in details such as materials, lighting, and colors. Ultimately, strong structural controllability is achieved in the industrial design rendering process, significantly improving design efficiency and rendering quality.

[0049] In some embodiments, the image rendering model includes an image rendering architecture and a conditional diffusion model architecture. The image rendering architecture includes multiple image rendering units connected in sequence, and the conditional diffusion model architecture includes injection units that correspond one-to-one with the multiple image rendering units. The image rendering units are used to generate rendering results, and the injection units are used to extract control features.

[0050] During the rendering process of the image rendering model, the first image rendering unit renders the received rendering result to obtain the first rendering result; the first injection unit corresponding to the first image rendering unit performs control feature extraction on the rendering result received by the first image rendering unit to obtain the first extraction result; the first rendering result is updated based on the weight parameters of the first injection unit and the first extraction result to obtain the updated first rendering result; the updated first rendering result is sent to the second image rendering unit as the rendering result received by the second image rendering unit; wherein, the first image rendering unit and the second image rendering unit are two adjacent image rendering units.

[0051] The image rendering architecture is the core component responsible for generating creative ideas; it can be understood as the backbone network of the generation process. It consists of multiple image rendering units stacked sequentially according to the processing flow. An image rendering unit can be a collection of one or more network layers, such as a ResNet block or a Transformer block. It receives the output from the previous unit and performs a core step of denoising and feature processing.

[0052] Meanwhile, the conditional diffusion model architecture is a bypass or control network specifically designed to process and transmit structural constraint information, and it also consists of multiple injection units. These injection units correspond one-to-one with the image rendering units in the image rendering architecture in terms of both number and hierarchy; that is, the first image rendering unit corresponds to the first injection unit, the second image rendering unit, and so on.

[0053] In this embodiment, the core function of the image rendering unit is to generate rendering results, that is, to denoise the input latent representation under the semantic guidance of the design text, and generate clearer features that include content and style. The core function of the injection unit is to extract control features. Based on the latent representation of the input line drawing control map, it extracts structured features at each level to guide the generation of the image at that level.

[0054] In some embodiments, the weight parameter is a learnable scaling factor (controlnet_conditioning_scale) that controls the strength of the injected structural constraint information.

[0055] In this embodiment, at each stage and feature level of iterative denoising, the backbone network (image rendering architecture) is simultaneously calibrated by a parallel bypass network (conditional diffusion model architecture) dedicated to structure creation. This method of continuously injecting structural information during processing effectively avoids the problem of structural information loss or distortion that may occur in deep networks due to information being passed layer by layer. From a data processing perspective, it ensures that the generated data strictly adheres to the constraints of the line drawing control chart at every level of abstraction, from low-level features (such as the precise position of lines) to high-level features (such as the overall outline and proportions of the product), thereby greatly enhancing the structural fidelity and accuracy of the final rendered image and achieving truly controllable generation.

[0056] In some embodiments, for the first rendering result h_1 output by the first image rendering unit, the first injection unit performs control feature extraction on the received rendering result to obtain the first extraction result Δh_l, where Δh_l is represented in residual form and s is the weight parameter, i.e., the learnable scaling factor (controlnet_conditioning_scale). Then, based on the weight parameter of the first injection unit and the first extraction result, the first rendering result is updated to obtain the updated first rendering result, which can be expressed as: h_l←h_l+s×Δh_l,0 <s≤1; The weight parameter takes values ​​in the range of (0, 1).

[0057] The first image rendering unit receives the rendering results as industrial design text and a low-dimensional latent space representation.

[0058] In some embodiments, the weight parameter takes values ​​in the range of (0.4, 1).

[0059] In some embodiments, the weight parameter is set to 0.8 by default.

[0060] In some embodiments, the image rendering architecture is a Transformer architecture, and the conditional diffusion model architecture is a ControlNet branch consistent with the image rendering architecture.

[0061] ControlNet is a conditional diffusion model architecture that achieves precise control over the generation process by replicating the neural network structure of the main diffusion model and adding an additional control layer.

[0062] In some embodiments, ControlNet branches are initialized using zero_modules, meaning that the weights and biases (i.e., the first extraction result Δh_l) of all injected units are initialized to zero. This allows the model to learn content generation in the early stages of training and then gradually learn how to fuse structural constraints, thus ensuring training stability. Through this additive fusion (also known as residual injection), the structural information of the line drawing is forcibly and losslessly injected into the features of the backbone network.

[0063] In some embodiments, each image rendering unit includes: a self-attention layer for integrating information from the received rendering results to obtain an integrated result; a cross-attention layer for aligning and fusing the integrated result with the industrial design text to obtain a fused result; and a feedforward neural network for processing the fused result to obtain the rendering result output by the image rendering unit.

[0064] In this embodiment, the role of the self-attention layer is to establish dependencies between different parts within the hidden state. Specifically, the self-attention mechanism calculates a correlation score between each element in the received rendering result (which can be understood as a local region or patch in the latent representation of the image) and all other elements. Based on this score, each element aggregates information from all other elements in a weighted manner. This process is equivalent to allowing each part of the image to see all other parts, thereby understanding the global context. Through this global information integration, the self-attention layer can ensure that the generated image is structurally consistent and physically plausible, and its output integration result is a more coherent rendering result optimized by the internal context.

[0065] The Cross-Attention Layer aligns and fuses the integrated result with the industrial design text, yielding a fused result. This layer is crucial for the model to understand and execute natural language instructions. It receives two inputs: the integrated result from the self-attention layer, which already contains rich internal contextual information; and a vector representation of the industrial design text processed by a specialized text encoder.

[0066] Specifically, the function of the cross-attention layer is to align and fuse these two different modalities (image information and text information). Mechanistically, it uses image information (the fusion result) as the query and text information as the key and value. Each region of the image (Query) queries the most relevant part of the text (Key) and extracts the semantic information (Value) of that part. For example, when processing the car body region, the query for that region will have a high correlation with the key "silver-gray metallic texture" in the design text, thus fusing the semantic value representing the silver-gray metallic texture into the fusion result for that region. In this way, the cross-attention layer accurately renders the style, material, color, and other attributes described in the text onto the image structure determined by the line art and the self-attention layer. The fusion result output by this process is an updated rendering result that maintains internal logical consistency while incorporating updated external semantic instructions.

[0067] A feed-forward network (FFN) is used to process the fusion result to obtain the rendering result output by the image rendering unit. A FFN typically consists of two linear transformation layers and a non-linear activation function (such as GELU). It receives the fusion result from the cross-attention layer and performs a pointwise, deep non-linear feature transformation on it. This processing step enables the model to learn more complex and abstract feature representations, enhancing the modeling capability of the entire image rendering unit. The output of the feed-forward network is the final rendering result output of that image rendering unit. It serves as the input to the next image rendering unit and is simultaneously fed to the corresponding injection unit for processing.

[0068] In this embodiment, the self-attention layer first ensures the inherent consistency of image generation in visual logic, resolving issues of global structure and lighting coordination. Subsequently, the cross-attention layer precisely injects the user's textual design intent into the image, achieving semantic control over artistic style and material details. Finally, the feedforward neural network further enhances the model's feature representation capabilities. This structure allows the model to simultaneously consider the internal harmony of the image and external instructions in each rendering step, thereby significantly improving the overall quality and controllability of the final generated rendering image.

[0069] In some embodiments, the self-attention layer employs learnable gating parameters to adjust the proportion of influence of the industrial design text and the low-dimensional latent space representation on the integration result; and / or The cross-attention layer uses learnable gating parameters to adjust the proportion of influence of the industrial design text and the low-dimensional latent space representation on the fusion result.

[0070] The cross-attention layer is responsible for fusing design text, while the self-attention layer is responsible for integrating internal visual information. However, in complex industrial design rendering scenarios, the degree to which different areas depend on external instructions (such as text) and internal context changes dynamically. For example, when rendering the metal body of a vehicle, the model needs to focus on the material descriptions in the design text, such as "brushed silver" or "mirror black"; while when rendering the ground shadows under the wheels, it should rely more on internal contextual information such as lighting and geometric position, where the material descriptions in the text may be irrelevant or even harmful interference. If a fixed, uniform information fusion method is used, it is easy to cause instruction leakage, such as incorrectly applying the metallic texture of the car body to the background or shadows, resulting in unnatural rendering results.

[0071] This embodiment introduces a gated attention fusion mechanism in the cross-attention layer and / or self-attention layer. The core idea of ​​this mechanism is to introduce a learnable gating parameter, which generates a dynamic gating value to intelligently control the flow and intensity of information, acting similarly to a valve that automatically adjusts the flow rate. The specific implementation of this mechanism can be expressed by the following formula: h←h+σ(α)⊙SiLU(W×c); h represents the hidden state, which is the integration result of the self-attention layer or the fusion result of the cross-attention layer; α represents the learnable gating parameter; σ represents the sigmoid function activation with an output range of (0, 1); W represents the projection matrix; c represents the conditional context, which is the design text of the industrial design; and ⊙ represents element-wise multiplication.

[0072] By introducing this gated attention fusion mechanism, the method in this embodiment achieves intelligent and differentiated control of multi-source information flows at the data processing level. It does not simply add textual and structural information together, but rather dynamically calculates the weights for information fusion using a learnable gating parameter α, and then precisely applies these weights to the conditional context c through element-wise multiplication ⊙. This approach allows the model to autonomously determine the intensity ratio of various guiding information fusions based on the specific needs of the generation task in different spatial locations and semantic regions. This ultimately results in a more harmonious overall rendering with more reasonable details, effectively avoiding conflicts and interference between different guiding information, thus achieving higher visual fidelity and artistic expression while adhering to macroscopic constraints.

[0073] In some embodiments, such as Figure 2 As shown, the rendering method also includes: Step 201: Determine the sharpening index corresponding to the updated first rendering result.

[0074] In some embodiments, the sharpness metric is a quantitative parameter used to measure the sharpness or detail of the current rendered result (a latent space representation). It can be determined in various ways; for example, one specific technical implementation involves calculating the variance of the Laplacian operator of the currently updated rendered result (a low-dimensional tensor) through a simplified, fast decoder or directly in the latent space. In image processing, Laplacian variance is a classic metric for measuring image focus and sharpness. A high variance value typically indicates the presence of numerous, rapidly changing edges and high-frequency details in the image, which, in the context of rendering, is a digital representation of oversharpening.

[0075] Step 202: Based on the sharpening index being greater than or equal to the index threshold, reduce the rescaling factor used by the image rendering unit.

[0076] The metric threshold is a pre-defined reference value that represents the upper limit of acceptable sharpness in the rendered result. This threshold can be determined empirically based on a large amount of experimental data to strike a balance between preserving necessary details and avoiding distortion.

[0077] The rescaling factor is a key normalized control parameter within the rendering unit. It is used to dynamically adjust the effective guidance scale generated by the design text during the rendering process. When the sharpening index determined in the above steps is greater than or equal to a preset index threshold, the system determines that the current rendering result shows signs of over-sharpening. At this time, the method will proactively reduce the value of the rescaling factor used by the image rendering unit.

[0078] Reducing the rescaling factor triggers or enhances the normalization correction for high-intensity guiding signals. Specifically, this normalization process subtracts a correction amount determined by the difference between the model's unconditional and conditional predictions from the current render generated by high guiding intensity. The underlying mechanism of reducing the rescaling factor is to increase the weight of this correction in the final result, effectively reversing overly sharp edges, oversaturated colors, and unnatural textures caused by overguiding towards a smoother, more natural state. If the sharpening metric does not reach a threshold, the rescaling factor can remain at its default value, with no or weak correction.

[0079] Through the dynamic normalization mechanism introduced in this embodiment, this method forms a precise internal feedback loop in the data processing flow. Instead of blindly and continuously applying strong guidance, it performs a quality assessment (determining sharpening metrics) at each processing step of the rendering unit. When the assessment results indicate a risk of distortion, it immediately adjusts the rendering parameters (reducing the rescaling factor) to enhance the smoothing and normalization of the current result. This step-by-step, real-time intervention effectively prevents the accumulation and amplification of sharpening distortion throughout the iterative denoising process. Ultimately, this mechanism ensures that even when using high guidance intensity to achieve a high degree of matching with the design text, the generated industrial design rendering still maintains delicate material textures and natural lighting transitions, avoiding common distortion problems and significantly improving the professional quality and visual realism of the final rendering.

[0080] In some embodiments, the formula used for normalization is: prediction=prediction-(prediction-model_pred)×(scale–1) / scale; Here, prediction is the preliminary prediction result, model_pred is the unconditionally generated result, and scale is the normalization coefficient, also known as the rescaling factor, with a default value of 1.0.

[0081] In some embodiments, the industrial design is for rail transit vehicles, and the engineering parameters include overall vehicle body parameters, front profile parameters, and component layout parameters.

[0082] like Figure 3 As shown, the initial line drawing corresponding to the industrial design is determined based on the engineering parameters of the industrial design, specifically including: Step 301: Based on the vehicle front profile parameters, calculate the vehicle front profile curve using cubic Bézier curves.

[0083] The cubic Bézier curve is a parametric curve widely used in computer graphics. It is uniquely determined by four control points (one start point, one end point, and the midpoint between the two control handles).

[0084] In some embodiments, the front profile parameters include the windshield position (1450mm from the front of the vehicle), the windshield tilt angle (44 degrees), and the position and dimensions of the coupler. The coupler position and dimensions include the coupler ellipse parameters. Then: The front profile curve is calculated using cubic Bézier curves, which must meet three constraints: the top must be tangent to the horizontal roofline, the middle section must pass through the windshield at a specified angle, and the bottom must be tangent to the coupler ellipse. The optimal connection point is calculated through iterative optimization to ensure the curve is smooth, continuous, and conforms to engineering specifications.

[0085] B(t) = (1-t) 3 P0+3(1-t) 2 tP1+3(1-t)t 2 P2+t 3 P3; Where t∈[0,1], B(t) is a cubic Bézier curve, and P0, P1, P2, and P3 are the control points of the Bézier curve, which are determined by solving the constraint equations to ensure that the tangents of the curve are continuous at the connection points.

[0086] Step 302: Determine the initial line drawing corresponding to the industrial design based on the front profile curve, overall vehicle parameters, and component layout parameters.

[0087] In some embodiments, the overall vehicle body parameters include defining the total length of the vehicle body (e.g., 26250mm), the height of the vehicle body section (e.g., 4050mm), the width of the vehicle body section (e.g., 3360mm), and the track gauge (e.g., 1435mm).

[0088] In some embodiments, the component layout parameters are defined as follows: the number of doors (e.g., 2), door width (1300mm), door height (1900mm), door starting position (4500mm from the front of the vehicle), and door spacing (16000mm); the number of windows (e.g., 8), window width (1400mm), window height (900mm), window starting position (6500mm from the front of the vehicle), and window spacing (1900mm); and bogie parameters, including wheelbase (2500mm), distance from the front of the vehicle, and height offset (-700mm).

[0089] In this embodiment, the generated front profile curve is used as the front end of the vehicle. Based on the vehicle length defined in the overall vehicle parameters, the endpoints of the front profile curve are extended or stretched along the vehicle length to form the side profile lines and top and bottom profile lines of the vehicle body. Finally, according to the precise coordinates defined in the component layout parameters, the boundary lines of components such as doors, windows, and lights are drawn on the already formed vehicle frame. Through this series of operations, a complete initial line drawing of the rail transit vehicle is finally generated, with all parts' dimensions and positions strictly conforming to the engineering parameters.

[0090] By introducing parametric Bézier curves to generate the crucial complex surface of the vehicle's front end, abstract contour parameters are transformed into mathematically precise smooth curve data. Compared to simple combinations of straight lines or arcs, this method generates a front end contour that is geometrically more accurate, visually smoother, and closer to the real needs of modern industrial design. When this high-quality front end contour is combined with other precise parametric vehicle body and component information to form the initial line drawing, it provides subsequent edge detection steps with input containing more effective structural details and fewer geometric imperfections. This allows the final generated line drawing control map to more faithfully reflect the designer's exquisite conception of the vehicle's streamlined aesthetics. When the image rendering model uses this line drawing control map for constraints, this high-precision structural information guides the model to generate rendering results with correct and smooth surface transitions, especially avoiding unnatural sharp angles or distortions in key curved surfaces such as the front end. This significantly improves the realism and professional quality of the final rendering, bringing the generated results to the standards of industrial applications.

[0091] In some embodiments, a front view of the vehicle's front end is drawn based on parameters such as the vehicle body cross-section width, the roof corner radius (e.g., 400mm), the bottom corner radius (e.g., 200mm), and the roof arc height (e.g., 150mm). The front view includes the vehicle body outline, the windshield (rounded rectangle), the windshield wipers, the vehicle hook (ellipse), and track representation.

[0092] In some embodiments, a side view of the vehicle body is drawn based on the overall length of the vehicle body, the front profile curve, the roofline, and the bottom line. The side view includes the front profile curve, the roofline, the bottom line, the rear vertical line, the side window (rounded rectangle), the door (rounded rectangle), the side wiper, and the bogie (rectangular frame + round wheel).

[0093] In some embodiments, such as Figure 4 As shown, after determining the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design, before performing edge detection on the initial line drawing and obtaining the line drawing control map, the following steps are also included: Step 401: Perform Gaussian filtering on the initial line drawing.

[0094] Gaussian filtering is a common image smoothing and denoising technique. In practice, initial line art generated by a program or imported from external sources may contain small, unwanted noise points or slight jagged edges. These noises numerically manifest as abrupt changes in pixel grayscale values. If subsequent edge detection is performed directly, they may be incorrectly identified as edges, introducing messy, meaningless lines into the final line art control image and interfering with the rendering model's structural understanding. By applying Gaussian filtering, this method uses a Gaussian kernel to convolve the initial line art image, effectively smoothing the image and suppressing these high-frequency noises, while having minimal impact on real, large-scale structural edges in the image. This lays the foundation for subsequent accurate gradient calculations.

[0095] Step 402: Use the Sobel operator to calculate the gradient magnitude and direction of the initial line drawing after Gaussian filtering; and perform non-maximum suppression processing based on the gradient magnitude and scheme.

[0096] The gradient is a vector describing the rate and direction of change of pixel intensity in an image. In this step, the Sobel operator, as a discrete first-order differential operator, is used to approximate the gradient of the image. Specifically, it convolves the smoothed image with two 3x3 convolution kernels in the horizontal and vertical directions, respectively, to obtain the horizontal gradient (Gx) and vertical gradient (Gy) of each pixel. Based on these two values, the gradient magnitude and gradient direction at that point can be further calculated. The gradient magnitude represents the drastic change in pixel value at that point; the larger the magnitude, the more likely the point is to be part of an edge. The gradient direction represents the direction of the fastest change in pixel value, which is usually perpendicular to the direction of the edge.

[0097] The calculated gradient magnitude map typically produces thick, bright bands to represent edges, rather than thin lines only one pixel wide. Non-maximum suppression (NMS) aims to refine these thick edges into lines only one pixel wide. Specifically, it iterates through each pixel in the gradient magnitude map, examining its gradient direction. Then, it compares the gradient magnitude of the current pixel with the gradient magnitudes of its two adjacent pixels along that direction. Only if the gradient magnitude of the current pixel is greater than its two adjacent pixels (i.e., the point is a local maximum) is its gradient magnitude retained; otherwise, the gradient magnitude is set to zero. This process ensures that only pixels with the most dramatic gradient changes along their edge directions are preserved, resulting in sharp, single-pixel edges.

[0098] Step 403: Based on the first threshold and the second threshold, determine the strong edges and weak edges in the initial line drawing after non-maximum suppression processing. The first threshold corresponds to the strong edges, and the second threshold corresponds to the weak edges.

[0099] Step 404: Preserve the strong edges and the parts where strong and weak edges connect.

[0100] In this step, while non-maximum suppression refines the edges, it may still retain some pseudo-edges caused by noise. This step aims to address this issue using a hysteresis thresholding method. Specifically, this method presets two thresholds: a higher first threshold and a lower second threshold. Then, based on the gradient magnitude of the pixels after non-maximum suppression, they are classified: points with a gradient magnitude greater than the first threshold are identified as strong edges, which are undoubtedly edge points; points with a gradient magnitude between the first and second thresholds are identified as weak edges, which may be part of a real edge or noise; points with a gradient magnitude less than the second threshold are directly suppressed and considered non-edges.

[0101] After classification, edge connectivity is performed. The core idea of ​​this process is that a true edge is usually continuous. Therefore, this method retains all strong edges and filters weak edges. Only weak edges that can be connected to at least one strong edge pixel through a path consisting of other weak or strong edge pixels are ultimately retained and promoted to true edges. All isolated weak edges that are not connected to any strong edges are suppressed.

[0102] Through the refined processing steps described above—Gaussian filtering, Sobel operator calculation, non-maximum suppression, and double-threshold hysteresis tracking—the line drawing control map generated in this embodiment exhibits significant advantages at the data level. Gaussian filtering and double-threshold detection effectively suppress interference from noise data; non-maximum suppression ensures that the data representing edges is refined into clear signals of single-pixel width; and hysteresis tracking reconnects edge data interrupted by noise or uneven lighting, guaranteeing the continuity and integrity of structural information. Ultimately, this high-quality line drawing control map provides an extremely accurate and robust structural constraint for the subsequent image rendering model, enabling the model to generate industrial design renderings with more accurate geometry, clearer details, and fewer artifacts, thereby further enhancing the overall technical effect of this invention.

[0103] In some embodiments, such as Figure 5 As shown, the design text and low-dimensional latent space representation of the industrial design are input into the image rendering model for rendering, and the rendering result output by the image rendering model is obtained, which also includes: Step 501: Receive the first input of the target display parameter among multiple sets of display parameters, including aspect ratio and resolution.

[0104] In some embodiments, the first input refers to one or more interactive operations performed by the user through the user interface. In a specific application scenario, the interface of the rendering software may provide a set of preset aspect ratio options (such as buttons or drop-down menus) and a text box for inputting specific resolution values. The user can first click the "16:9" button, and then type "1920" and "1080" in the resolution input box. This series of interactive operations constitutes the first input, the purpose of which is to specify a target display parameter, that is, the display specifications that the user expects the final rendered image to have.

[0105] Step 502: In response to the first input, determine the target display parameters corresponding to the rendering result output by the image rendering model.

[0106] In this embodiment, user control over output specifications is brought forward to before the rendering process begins, ensuring that all data processing in the image generation workflow is performed natively within a latent space that conforms to the target size. This method dynamically adjusts the dimensions of the initial noise latent representation by receiving user input, thereby ensuring that all calculations in the image rendering model, when fusing semantic information from the design text and structural information from the line drawing control diagram, are completed within an internal space proportional to the final output size. This approach fundamentally avoids the need for cropping or scaling the image after rendering, as these post-processing operations often lead to image quality degradation, loss of detail, or proportional distortion. Finally, when the rendered target result is fed into the VAE decoder, it naturally decodes and generates a rendered image that perfectly matches the aspect ratio and resolution specified by the user. This greatly enhances the practicality and flexibility of the generation method, enabling it to directly meet the precise image specification requirements of different application scenarios (such as high-definition printing, web display, presentations, etc.).

[0107] In some embodiments, the multiple sets of display parameters can be at least two of the following: 1:1 (1024×1024) scale is suitable for social media postings and solution thumbnails; 16:9 (1664×928) is suitable for landscape displays and poster design; 4:3 (1472×1104) is suitable for traditional documents and PPT illustrations; 4:1 (2560×448) is suitable for multi-view display of industrial products (front view + side view).

[0108] In some embodiments, a rendering graph generation system is proposed to perform a rendering graph generation method. Specifically, the rendering graph generation system includes: Parametric line art generation module: Based on user-input engineering parameters (total body length 26250mm, section height 4050mm, number of doors 2, number of windows 8, etc.), it uses cubic Bézier curves to calculate the front profile curve (satisfying three constraints: roof tangency, windshield position, and coupler ellipse tangency), automatically generating accurate structural line art (i.e., initial line art). Supports multiple output formats including PNG, SVG, and transparent PNG.

[0109] Canny Edge Detection Module: Extracts the control map of the line drawing from the structural line drawing using the Canny edge detection algorithm. It employs steps such as Gaussian smoothing, Sobel gradient calculation, non-maximum suppression, and double threshold detection to ensure that the extracted edges are continuous and accurate.

[0110] VAE Encoding Module: Encodes the line art control graph to latent space.

[0111] The ControlNet module employs a zero-initialization ControlNet branch to extract structural constraint features from the line drawing control graph. It dynamically adjusts the line drawing constraint strength using a learnable scaling factor (default 0.8, range 0.4-1.0), injecting line drawing features as residuals into each Transformer layer during the denoising process, thus achieving dynamic fusion of line drawing constraints and image generation.

[0112] Gated attention fusion module: This module controls the fusion intensity of conditional information through learnable gating parameters. The sigmoid activation function is used to restrict the gating parameters to the range of (0, 1), adaptively adjusting the contribution ratio of textual and visual conditions to achieve progressive fusion of multiple conditions.

[0113] This system overcomes the bottlenecks of insufficient structural controllability and low line drawing efficiency of traditional diffusion models by using parametric line drawing generation, precise control of line drawing constraints, and multi-condition fusion mechanisms, providing a complete technical path for industrial design visualization.

[0114] Table 1 shows the modules, their corresponding functions, and the technologies used in the rendering graph generation system.

[0115] Table 1

[0116] As shown in Table 1, the rendering image generation system also includes an aspect ratio adaptation module and a TrueCFG normalization module to achieve adaptive resolution adjustment and classifier-free guidance (CFG) amplitude normalization.

[0117] In one specific embodiment, the engineering parameters are: • Overall vehicle length: 26250mm; • Vehicle body section height: 4050mm, width: 3360mm; • Doors: 2, width 1300mm, height 1900mm, starting position 4500mm; • 8 windows, 1400mm wide, 900mm high, starting position 6500mm; • Cue words (i.e., design text for industrial design): Multi-perspective_v1, a silver-gray high-speed train, red auspicious cloud pattern decorative strip, technological style; • ControlNet strength (i.e., weight parameter): 0.8.

[0118] The processing flow is as follows: 1. Parametric line drawing generation: Based on the input parameters, the outline of the vehicle front is calculated using Bézier curves to generate a structural line drawing that includes a front view and a side view.

[0119] 2. Canny edge detection: Extract line art edge features as ControlNet control conditions.

[0120] 3. ControlNet Generation: Generates rendered images based on line drawing constraints, ensuring accurate vehicle proportions and door / window positions.

[0121] Output result: • Precise structural line drawings (front view + side view).

[0122] • High-quality rendered images that conform to line art constraints.

[0123] • Silver-gray body with red auspicious cloud paint.

[0124] In this embodiment, structural line drawings are automatically generated based on engineering parameters using a parametric approach, eliminating the need for manual drawing and reducing line drawing generation time from hours to seconds. Bezier curve calculations ensure the engineering rationality of the vehicle's front profile. The ControlNet mechanism enables precise injection of line drawing constraints, resulting in a generation that highly matches the design intent in key structural elements such as vehicle proportions and door / window positions, avoiding the proportional imbalance issues common in pure text generation. TrueCFG normalization effectively suppresses over-sharpening, generating natural and detailed images. A gating attention mechanism balances the fusion of multiple conditions, preventing any single condition from becoming overly dominant.

[0125] The rendering image generation apparatus provided by the present invention will be described below. The rendering image generation apparatus described below and the rendering image generation method described above can be referred to in correspondence.

[0126] In some embodiments, such as Figure 6 As shown, a rendering graph generation apparatus is proposed, comprising: The processing module 601 is used to determine the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design, and to perform edge detection on the initial line drawing to obtain the line drawing control map. Encoding module 602 is used to input the line drawing control diagram into the variational autoencoder for encoding to obtain a low-dimensional latent space representation; The rendering module 603 is used to input the industrial design text and the low-dimensional latent space representation into the image rendering model for rendering, and obtain the target rendering result output by the image rendering model. The target rendering result is generated based on the industrial design text under the constraint of the low-dimensional latent space representation. The decoding module 604 is used to input the target rendering result into the variational autoencoder for decoding to obtain the rendering image of the industrial design.

[0127] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a rendering image generation method, which includes: determining the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design, and performing edge detection on the initial line drawing to obtain a line drawing control map; inputting the line drawing control map into a variational autoencoder for encoding to obtain a low-dimensional latent space representation; inputting the design text of the industrial design and the low-dimensional latent space representation into an image rendering model for rendering to obtain a target rendering result output by the image rendering model, wherein the target rendering result is generated based on the design text of the industrial design under the constraints of the low-dimensional latent space representation; and inputting the target rendering result into a variational autoencoder for decoding to obtain a rendering image of the industrial design.

[0128] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0129] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the rendering image generation method provided by the above methods. The method includes: determining the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design, and performing edge detection on the initial line drawing to obtain a line drawing control map; inputting the line drawing control map into a variational autoencoder for encoding to obtain a low-dimensional latent space representation; inputting the design text of the industrial design and the low-dimensional latent space representation into an image rendering model for rendering to obtain a target rendering result output by the image rendering model, wherein the target rendering result is generated based on the design text of the industrial design under the constraints of the low-dimensional latent space representation; and inputting the target rendering result into a variational autoencoder for decoding to obtain a rendering image of the industrial design.

[0130] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the rendering image generation method provided by the above methods. The method includes: determining an initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design, and performing edge detection on the initial line drawing to obtain a line drawing control map; inputting the line drawing control map into a variational autoencoder for encoding to obtain a low-dimensional latent space representation; inputting the design text of the industrial design and the low-dimensional latent space representation into an image rendering model for rendering to obtain a target rendering result output by the image rendering model, wherein the target rendering result is generated based on the design text of the industrial design under the constraints of the low-dimensional latent space representation; and inputting the target rendering result into a variational autoencoder for decoding to obtain a rendering image of the industrial design.

[0131] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0132] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating a rendered image, characterized in that, include: Based on the engineering parameters of the industrial design, the initial line drawing corresponding to the industrial design is determined, and edge detection is performed on the initial line drawing to obtain the line drawing control map; The line drawing control diagram is input into a variational autoencoder for encoding to obtain a low-dimensional latent space representation. The industrial design text and the low-dimensional latent space representation are input into the image rendering model for rendering to obtain the target rendering result output by the image rendering model. The target rendering result is generated based on the industrial design text under the constraints of the low-dimensional latent space representation. The target rendering result is input into the variational autoencoder for decoding to obtain the rendering image of the industrial design.

2. The rendering map generation method of claim 1, wherein, The image rendering model includes an image rendering architecture and a conditional diffusion model architecture. The image rendering architecture includes multiple image rendering units connected in sequence. The conditional diffusion model architecture includes injection units that correspond one-to-one with the multiple image rendering units. The image rendering units are used to generate rendering results. The injection units are used to extract control features. During the rendering process of the image rendering model, the first image rendering unit renders the received rendering result to obtain the first rendering result; The first injection unit corresponding to the first image rendering unit performs control feature extraction on the rendering result received by the first image rendering unit to obtain a first extraction result; The first rendering result is updated based on the weight parameters of the first injection unit and the first extraction result to obtain the updated first rendering result; The updated first rendering result is sent to the second image rendering unit as the rendering result received by the second image rendering unit; The first image rendering unit and the second image rendering unit are two adjacent image rendering units.

3. The rendering image generation method according to claim 2, characterized in that, Each of the image rendering units includes: The self-attention layer is used to integrate information from the received rendering results to obtain the integrated result. A cross-attention layer is used to align and merge the integration result and the industrial design text to obtain a fusion result; A feedforward neural network is used to process the fusion result to obtain the rendering result output by the image rendering unit.

4. The rendering image generation method according to claim 3, characterized in that, The self-attention layer uses learnable gating parameters to adjust the proportion of influence of the industrial design text and the low-dimensional latent space representation on the integration result; and / or The cross-attention layer uses learnable gating parameters to adjust the proportion of influence of the industrial design text and the low-dimensional latent space representation on the fusion result.

5. The rendering image generation method according to claim 3, characterized in that, The method for generating the rendered image also includes: Determine the sharpening index corresponding to the updated first rendering result; Based on the sharpening index being greater than or equal to the index threshold, the rescaling factor used by the image rendering unit is reduced.

6. The method for generating a rendered image according to any one of claims 1 to 5, characterized in that, The industrial design is for rail transit vehicles, and the engineering parameters include overall vehicle body parameters, front profile parameters, and component layout parameters. The determination of the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design specifically includes: Based on the aforementioned front profile parameters, the front profile curve is calculated using cubic Bézier curves. The initial line drawing corresponding to the industrial design is determined based on the vehicle front profile curve, the overall vehicle body parameters, and the component layout parameters.

7. The method for generating a rendered image according to any one of claims 1 to 5, characterized in that, After determining the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design, and before performing edge detection on the initial line drawing to obtain the line drawing control map, the method further includes: The initial line drawing is subjected to Gaussian filtering; The Sobel operator is used to calculate the gradient magnitude and direction of the initial line drawing after Gaussian filtering; and non-maximum suppression is performed based on the gradient magnitude and the scheme. Based on a first threshold and a second threshold, strong edges and weak edges in the initial line drawing after the non-maximum suppression processing are determined, wherein the first threshold corresponds to the strong edges and the second threshold corresponds to the weak edges; The strong edge and the portion where the strong edge connects to the weak edge are retained.

8. The method for generating a rendered image according to any one of claims 1 to 5, characterized in that, The step of inputting the industrial design text and the low-dimensional latent space representation into an image rendering model for rendering, and obtaining the rendering result output by the image rendering model, further includes: Receive a first input of a target display parameter from a plurality of display parameters, the display parameters including aspect ratio and resolution; In response to the first input, the target display parameters corresponding to the rendering result output by the image rendering model are determined.

9. A rendering image generation apparatus, characterized in that, include: The processing module is used to determine the initial line drawing corresponding to the industrial design based on the engineering parameters of the industrial design, and to perform edge detection on the initial line drawing to obtain the line drawing control map; The encoding module is used to input the line drawing control diagram into the variational autoencoder for encoding to obtain a low-dimensional latent space representation; The rendering module is used to input the design text of the industrial design and the low-dimensional latent space representation into the image rendering model for rendering, and to obtain the target rendering result output by the image rendering model. The target rendering result is generated based on the design text of the industrial design under the constraints of the low-dimensional latent space representation. The decoding module is used to input the target rendering result into the variational autoencoder for decoding to obtain the rendering image of the industrial design.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the rendering graph generation method as described in any one of claims 1 to 8.

Images