Method and device for modifying producibility of a design of a figurine, product, medium

Abstract

A method, apparatus, product, and medium for manufacturability correction of a trendy toy design drawing are disclosed. The manufacturability correction method includes: receiving the original design drawing and material parameters of the trendy toy; calling a pre-trained multimodal large model to segment the original design drawing into regions, analyzing the multiple image regions obtained after segmentation, and identifying defective regions in the multiple image regions based on a pre-set material and process knowledge base; calling the multimodal large model to obtain modification prompts based on the process constraint rules violated by the defective regions; and calling a pre-trained image generation model to redraw the defective regions belonging to the modification area based on the modification prompts, thereby obtaining the modified design drawing. Using the above scheme, targeted redrawing of defective regions in the modification area can be achieved, ensuring the preservation of the core image of the trendy toy while correcting the defective regions, thus improving the efficiency and reliability of manufacturability correction.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, apparatus, product, and medium for manufacturability modification of trendy toy design drawings. Background Technology

[0002] In the process of designing and mass-producing trendy toys (vinyl, plush, resin figurines, acrylic products, etc.), manufacturability modifications are often required. This is because trendy toy designers focus on visual creativity and often lack knowledge of manufacturing processes. Before mass production, the design drawings must be modified to address manufacturing constraints (such as demolding feasibility, wall thickness uniformity, and detail sharpness). Manufacturability modification refers to adjusting parts of the design drawings that do not meet manufacturing constraints (such as difficulty in demolding, structural fragility, and unstable center of gravity) to meet mass production requirements. Current manufacturability modification processes rely on process engineers reviewing and designers repeatedly communicating and revising drawings, resulting in long iteration cycles and high costs.

[0003] Those skilled in the art have attempted to use general-purpose multimodal large models or existing artificial intelligence image generation tools to achieve manufacturability corrections, but the reliability of the correction results is poor. Firstly, each collectible toy has an associated IP, and the core value of a collectible toy lies in the recognizability of the IP, while general-purpose models often change IP elements when correcting defects; secondly, different materials have drastically different process requirements for the same design feature, and existing tools are difficult to dynamically adjust and modify. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide at least one method for manufacturability correction of trendy toy design drawings. By dividing the design drawings into a locked area and a modified area, and combining this method with a material and process knowledge base to identify defective areas, the method enables targeted redrawing of defective areas in the modified area. This ensures the preservation of the core image of the trendy toy while correcting defective areas, thereby improving the efficiency and reliability of manufacturability correction.

[0005] In a first aspect, the present invention provides a method for manufacturability modification of a trendy toy design, the method comprising:

[0006] Receive the original design drawings and material parameters of trendy toys;

[0007] The original design drawing is segmented by calling a pre-trained multimodal large model, and the resulting image regions are analyzed. The image regions are divided into at least modification and locking regions based on a preset image feature classification rule. Defect regions are identified in the image regions based on a preset material and process knowledge base, which contains process constraint rules for various materials. The defect region refers to the image region in the image regions that violates the process constraint rules corresponding to the material parameters.

[0008] The multimodal large model is invoked, and modification prompts are obtained based on the process constraint rules violated by the defect region;

[0009] The pre-trained image generation model is invoked, and the defective areas belonging to the modification area are redrawn based on the modification prompt words to obtain the modified design drawing;

[0010] in,

[0011] The preset image feature classification rule divides the multiple image regions into at least a modification region and a locked region, including:

[0012] The saliency of each image region is determined based on a pre-defined saliency detection algorithm;

[0013] Image regions with a significance level higher than the first threshold are identified as candidate locking regions;

[0014] Curvature analysis is performed on multiple lines contained in the candidate locking region. If the rate of change of curvature of any line exceeds a second threshold, the candidate locking region is determined as the locking region.

[0015] Optionally, the method for modifying the manufacturability of the toy design drawings further includes:

[0016] For the modified design drawing, the image feature loss gradient and manufacturability loss gradient are calculated for the defect area belonging to the modification area based on the preset loss function. The image feature loss gradient reflects the similarity between the modified design drawing and the original design drawing, and the manufacturability loss gradient reflects the degree to which the modified design drawing meets the process constraint rules corresponding to the material parameters. The loss function is pre-constructed based on the process constraint rules.

[0017] Calculate or obtain the weight coefficients of the image feature loss gradient and the productivity loss gradient, and perform a weighted calculation based on the image feature loss gradient and the productivity loss gradient to obtain the total gradient;

[0018] The image generation model is invoked, and the modified design drawing is modified based on the total gradient.

[0019] Optionally, the method for modifying the manufacturability of the toy design drawings further includes:

[0020] Repeat the following steps until the preset iteration termination condition is met:

[0021] Receive the design drawing after the last modification, and update the image feature loss gradient and the productiveization loss gradient based on the preset loss function;

[0022] The updated total gradient is obtained by weighting the updated image feature loss gradient with the updated productivity loss gradient.

[0023] The image generation model is invoked, and the previously modified design drawing is modified based on the updated total gradient.

[0024] Optionally, the preset iteration termination condition includes at least one of the following:

[0025] The number of iterations has reached the preset maximum.

[0026] The manufacturability loss value is less than a preset process tolerance threshold, and the manufacturability loss value is calculated based on the loss function.

[0027] Optionally, when performing weighted calculations, the weight coefficients of the image feature loss gradient and the productability loss gradient are preset fixed values, wherein the weight coefficient of the image feature loss gradient is greater than the weight coefficient of the productability loss gradient.

[0028] Optionally, as the number of iterations increases, the weight coefficient of the image feature loss gradient remains unchanged or increases, while the weight coefficient of the productivity loss gradient decreases.

[0029] Optionally, before performing the weighted calculation based on the updated image feature loss gradient and the updated productivity loss gradient, the method further includes:

[0030] Calculate the angle between the directions of the image feature loss gradient and the productivity loss gradient;

[0031] If the included angle is greater than a preset angle threshold, then the weight coefficient of the productivity loss gradient is reduced.

[0032] Optionally, the modification area is divided into at least a first-level modification area and a second-level modification area according to the preset image feature classification rule. The first-level modification area and the second-level modification area are assigned different weight coefficients of the image feature loss gradient, and the weight coefficient of the image feature loss gradient of the first-level modification area is greater than the weight coefficient of the image feature loss gradient of the second-level modification area.

[0033] Optionally, the identification of defect regions among the multiple image regions based on a preset material and process knowledge base includes:

[0034] The multimodal large model is invoked to perform feature measurements on each image region, generating feature data;

[0035] The feature data is compared with the constraint range in the process constraint rules corresponding to the material parameters;

[0036] If the feature data is outside the constraint range, the image region to which the feature data belongs is identified as a defect region, and the deviation value of the feature data from the corresponding constraint range is recorded.

[0037] Secondly, the present invention provides a manufacturability modification device for a trendy toy design drawing, comprising:

[0038] The data receiving unit is used to receive the original design drawings and material parameters of the trendy toy;

[0039] The multimodal analysis unit is used to call a pre-trained multimodal large model to perform region segmentation on the original design drawing, and to analyze the multiple image regions obtained after segmentation. Specifically, based on a preset image feature grading rule, the multiple image regions are divided into at least a modification region and a locked region. Based on a preset material and process knowledge base, defect regions in the multiple image regions are identified. The material and process knowledge base contains process constraint rules for various materials. The defect region refers to the image region in the multiple image regions that violates the process constraint rules corresponding to the material parameters.

[0040] The prompt word generation unit is used to call the multimodal large model and obtain modification prompt words based on the process constraint rules violated by the defect area;

[0041] The image correction unit is used to call a pre-trained image generation model and redraw the defective areas belonging to the modification area based on the modification prompt words to obtain the modified design drawing.

[0042] The step of dividing the multiple image regions into at least modification regions and lock regions based on preset image feature grading rules includes: determining the saliency of each image region based on a preset saliency detection algorithm; identifying image regions with saliency higher than a first threshold as candidate lock regions; performing curvature analysis on multiple lines contained in the candidate lock regions, and if the rate of change of curvature of any line exceeds a second threshold, then the candidate lock region is identified as the lock region.

[0043] Thirdly, the present invention also provides a computer-readable storage medium, which is a non-volatile storage medium or a non-transient storage medium, on which a computer program is stored, wherein the computer program is executed by a computer to perform any of the above-described methods for manufacturability modification of trendy toy design drawings.

[0044] Fourthly, the present invention also provides a computer program product, including a computer program / instruction, wherein when the computer program / instruction is run by a computer, the manufacturability modification method of any of the above-described trendy toy design drawings is executed.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] This invention divides the design into locked and modified areas using a preset hierarchical rule based on saliency detection and curvature analysis. This allows for the identification of regions containing core IP characteristics as unmodifiable locked areas, enabling differentiated processing of different image regions in the design drawing. Locked areas are not allowed to be modified, while defective areas in the modified area are redrawn, preventing IP image distortion due to excessive modifications. By combining a material and process knowledge base to identify defective areas and generate modification prompts, the image generation model can specifically correct image regions that violate process constraints, effectively improving the manufacturability of the design drawing while maintaining the original IP image. Because the locked areas are strictly protected, even if the modification prompts guide significant geometric adjustments during redrawing, the recognizable features of the IP image will not be affected, ensuring the reliability of the correction results.

[0047] Furthermore, by calculating the gradient of image feature loss and the gradient of manufacturability loss and performing a weighted summation, the total gradient guides the secondary correction of the design drawing. This finds a better compromise between retaining core image features and meeting process constraints, effectively solving the problem of overshoot or insufficient correction that may occur in a single redraw.

[0048] Furthermore, the weight coefficient of the image feature loss gradient is preset to be greater than the weight coefficient of the manufacturability loss gradient, so that the total gradient will be biased towards maintaining the image features, thereby suppressing excessive process modifications at the expense of the core IP image.

[0049] Furthermore, by calculating the angle between the directions of the image feature loss gradient and the manufacturability loss gradient, it is possible to identify and address situations where there is a conflict between the two. When preserving IP features conflicts with satisfying process constraints, the guiding strength of the process constraints is weakened, prioritizing the protection of the IP image from damage and avoiding the distortion of core features caused by blindly following process rules. Attached Figure Description

[0050] Figure 1This is a flowchart of a method for manufacturability modification of a trendy toy design drawing according to an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of a manufacturability modification device for a trendy toy design drawing according to an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. 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.

[0053] As mentioned in the background section, existing general-purpose models, whether multimodal large models or image generation models, cannot simultaneously ensure manufacturability and maintain the IP image when processing design drawings. This is the core challenge in modifying the manufacturability of trendy toys. It is necessary to accurately identify and retain the core identifying features that constitute the IP image, while also detecting defects that violate the constraints of specific materials and processes and making targeted modifications. Otherwise, even if it is modified into a mass-producible product, it will lose its IP and have no meaning in manufacturing.

[0054] The inventors further analyzed that different areas of a toy's design do not have equal value for the IP image. Some image areas (such as facial features, iconic outlines, and typical accessories) constitute the core identifying characteristics of the IP, and changing them will render the character unrecognizable. Other image areas (such as limbs, color block boundaries, and base backgrounds) may require adjustment due to violations of specific material manufacturing constraints; even significant modifications will not affect the IP's recognizability. Therefore, the core solution lies in constructing a method that can distinguish between image areas that "must be absolutely locked" and those that "can be engineered for modification." Image areas that "must be absolutely locked," i.e., locked areas, cannot be modified in subsequent revisions to ensure the core IP image remains unchanged and the core value of the toy is not compromised. Image areas that "can be engineered for modification," i.e., modified areas, allow for modifications in a production-ready direction even if manufacturing constraints are violated, enabling mass production of the product.

[0055] To address this technical problem, this invention provides a method for manufacturability modification of trendy toy design drawings. The method utilizes a multimodal large model to segment the original design drawings into regions and, based on preset image feature grading rules, specifically including saliency detection and curvature analysis, divides each image region into a locked region (i.e., an image region that cannot be modified) and a modified region (i.e., an image region that allows engineering adjustments).

[0056] Specifically, saliency detection is used to locate visual focal areas (such as faces and iconic accessories), while curvature analysis is used to identify feature lines with unique geometric contours (such as tufts of hair, pointed ears, and canine teeth). The combination of these two methods ensures that the locked area accurately covers the core features of the IP. Furthermore, the multimodal large model generates modification prompts in natural language based on the specific rules violated by the defective area. Then, the image generation model is called to redraw only the defective areas belonging to the modification area, while strictly maintaining the pixel integrity of the locked area.

[0057] The above methods not only preserve the core identifying features of trendy toy IPs, but also automatically eliminate defects that violate material and process constraints, significantly improving the efficiency and reliability of design drawings for productive modification.

[0058] To make the above-mentioned objectives, features and beneficial effects of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0059] The manufacturability modification method for trendy toy design drawings provided in this embodiment of the invention can be widely applied to the digital development and manufacturing of various derivative products such as trendy toys, figurines, and blind boxes, for example, in an enterprise manufacturing execution system (MES, hereinafter referred to as the system). This system can be a cloud-based manufacturing collaboration platform. The backend of the system can run on a cloud server for processing and storing data; the frontend can be implemented based on a webpage or app for inputting design drawings and viewing various data processing results.

[0060] like Figure 1 As shown, Figure 1 This is a flowchart of a method for manufacturability modification of a trendy toy design drawing according to the first embodiment of the present invention.

[0061] Specifically, in this embodiment, the method for modifying the manufacturability of trendy toy design drawings includes the following steps:

[0062] Step S101 receives the original design drawings and material parameters of the trendy toy.

[0063] Specifically, the original design drawing can be a user-uploaded 2D concept drawing, 3D rendering, or hand-drawn sketch. Material parameters are selected or specified by the user, such as slush molding, PVC soft rubber, resin, plush, or acrylic. For materials like plush that don't require demolding, the material parameters include the material selection and image scale, such as 10 pixels = 1 mm. For other materials requiring demolding, such as injection molding or slush molding, the material parameters further include the specified mold opening direction. This direction serves as the benchmark for measuring other process parameters, the definitions of which will be explained in detail below. In practical applications, this mold opening direction is usually the viewing direction of the design drawing. For example, when the design drawing is a front view or side view, the mold opening direction is usually specified as the corresponding orthographic or side projection viewing direction.

[0064] Among them, process parameters such as "demolding angle," "wall thickness," and "minimum line width" are all geometric measurements within the plane of the two-dimensional image under a specified viewing angle. Taking the front view as an example, the original design drawing is a two-dimensional rasterized image of the toy under the orthographic projection view corresponding to the mold opening direction. This view is pre-specified to be consistent with the mold opening direction. "Demolding angle" refers to the angle between the surface of the object's sidewall and the mold opening direction. In the orthographic projection view of this embodiment, it is defined as: the angle between the tangent at a point on the contour line and the vertical direction of the image (defined as the positive Y-axis direction in the front view), taking an acute angle within the range of 0° to 90° (where 0° indicates that the tangent is parallel to the mold opening direction, making demolding the most difficult; 90° indicates that the tangent is perpendicular to the mold opening direction, making demolding the easiest), which can be directly measured. "Wall thickness" is defined as: for closed areas in the design drawing (such as arms, torsos, etc.), the wall thickness is defined as the vertical distance between the edges of the outer contour. In actual calculations, the boundary contour of the closed region is first extracted. Then, the minimum value of all vertical distances to both sides within the contour is taken, and the physical wall thickness is obtained after conversion using a scale. The "minimum line width" is defined as: for slender lines in the design drawing (such as clothing patterns, hair strands, etc.), the minimum line width is defined as the shortest vertical distance between the two edges of the line, which is converted using a scale to obtain the physical line width. In specific embodiments, this is also referred to as the actual line width.

[0065] Step S102: Call the pre-trained multimodal large model to perform region segmentation on the original design drawing, analyze the multiple image regions obtained after segmentation, and identify the defect regions in the multiple image regions based on the pre-set material and process knowledge base.

[0066] Specifically, region segmentation can employ deep learning-based semantic segmentation or instance segmentation methods to divide different objects or components in the design drawing into independent image regions. For example, the head, torso, arms, legs, accessories, etc. of a character can be segmented into different regions.

[0067] In one embodiment, a pre-trained semantic segmentation network (such as U-Net, DeepLabV3+, or SegFormer) is integrated or invoked within the multimodal large model. This network takes the original design image as input and outputs a semantic segmentation map of the same size as the original design image. Each pixel in the semantic segmentation map has a value between 0 and N-1, where N is the preset number of categories. Each integer value represents a semantic category: 0 represents "background," 1 represents "head," 2 represents "arm," 3 represents "accessory," and so on. Based on this semantic segmentation map, the system extracts consecutive pixel sets with the same category label into an image region and attaches a semantic label to each region, such as "background," "head," "arm," "accessory," etc.

[0068] In another embodiment, to more precisely distinguish different instances of the same category, such as differentiating between the left and right eyes or multiple fingers, the system employs an instance segmentation network (such as Mask R-CNN). The instance segmentation network outputs an independent mask and corresponding category label for each instance. The system processes these instance masks as independent image regions.

[0069] Furthermore, based on preset image feature grading rules, the multimodal large model divides multiple image regions into at least modification regions and lock regions. Specifically, the system determines the saliency of each image region based on a preset saliency detection algorithm. The saliency detection algorithm can use a pre-trained deep learning model (such as U2-Net, BASNet, etc.), taking the original design image as input and outputting a saliency heatmap of the same size as the original design image. The value of each pixel is between 0 and 1, indicating the degree to which the pixel attracts attention. For each segmented image region, the system calculates the average or maximum saliency score of all pixels in that region as the saliency score of that region. The system identifies image regions with saliency higher than a first threshold as candidate lock regions, while those with saliency less than or equal to the first threshold are classified as modification regions. The first threshold can be set empirically, for example, 0.7, to filter out the most visually striking regions, which typically correspond to the character's facial features, iconic accessories, or core symbols.

[0070] In practice, to more accurately determine the locking area, the system performs curvature analysis on multiple lines contained within the candidate locking area.

[0071] Specifically, the system calculates the curvature of each point on every line within the candidate locking area, thus obtaining the rate of change of curvature along the line. The rate of change of curvature reflects the degree of change in the line's bending; for example, the rate of change of curvature is greater at sharp corners and turning points. If the rate of change of curvature of any line exceeds a second threshold, it indicates that the candidate locking area has unique geometric features and is a core identifying element of the trendy toy image; the system then designates this candidate locking area as a locked area. Conversely, if the rate of change of curvature of all lines within the candidate locking area does not exceed the second threshold, even if the area has high saliency, it is still designated as a modification area to avoid over-protection affecting manufacturability process adjustments.

[0072] Here, the lines refer to the edge lines of the outer contour of the trendy toy extracted from the segmented image region, used to characterize the geometric shape and feature orientation of the component. These contour edge lines are extracted from the segmented image region using conventional edge detection algorithms (e.g., the Canny edge detection algorithm), resulting in closed or open contour lines composed of continuous ordered pixels. This is a conventional technique in the field and will not be elaborated further here.

[0073] The process of curvature analysis for candidate locking areas is as follows: Continuous sampling points are taken along the extracted contour edge line, and the curvature of each sampling point is calculated using geometric methods. Curvature is used to characterize the degree of bending of the line at that point. Further calculation of the curvature change amplitude between adjacent sampling points yields the curvature change rate. The larger the curvature change rate, the more prominent the geometric features of the line and the higher its contribution to the IP image's recognizability. In one specific implementation, the system uses the included angle of three consecutive points as a discrete geometric estimate of curvature: Three consecutive points A, B, and C are taken sequentially on the contour line, and the angle θ between vectors BA and BC is used as a discrete representation of the curvature at point B (the smaller the angle, the sharper the bend at point B, and the greater the corresponding curvature). Points are taken along the line, resulting in a series of included angle values. The difference between two adjacent sets of included angles is a discrete approximate estimate of the curvature change rate. Therefore, the "angle change" mentioned below corresponds to the "curvature change rate" referred to in this paragraph.

[0074] The second threshold can be set based on experience in trendy toy design, and is generally set between 10 and 30 degrees. Preferably, 15 degrees is used as an example for the following explanation.

[0075] Specifically, let's take the outer contours of the upper arm and hand of a collectible toy character as examples to further illustrate this. For a relatively smooth arm contour, the line is a smooth arc from the shoulder to the wrist. Take three consecutive points A, B, and C on this line and calculate the angle θ between vectors BA and BC. Typically, on a smooth arc, the angle formed by any three adjacent points is close to 160 to 180 degrees. Move along the upper arm line and take three more consecutive points, calculating the angle using the same method. For example, if the angle at one point is 175 degrees and at the next is 170 degrees, the change is 5 degrees. The second threshold is 15 degrees; 5 degrees is less than the threshold, so the system determines that there is no sharp turn in the upper arm and it does not belong to the core recognition features. Therefore, this area will not be locked, allowing for subsequent adjustments based on process requirements.

[0076] As for the outline of the claw, it extends from the base of the fingers to the fingertips, forming a sharp angle at the fingertips. Three consecutive points are taken on the line: A at the base of the fingers, B at the fingertip, and C slightly behind the fingertip. At this point, the angle θ between vectors BA and BC might only be 30 degrees. Further, three adjacent points are taken: A' slightly in front of the fingertip, B' slightly behind the fingertip, and C' some distance behind the fingertip. The calculated angle might be 40 degrees. The change in angle between these two points is 10 degrees. Moving forward a step, the fingertip is a smooth segment (three points are taken on the fingertip segment, with a calculated angle of 170 degrees). The transition from the fingertip (170 degrees) to the sharp fingertip segment (40 degrees) represents a change of 130 degrees, far exceeding the 15-degree threshold. The system determines that this point represents a sharp transition, a core identifying feature of the toy character (such as sharp claws). Therefore, this area is designated as a locked area and will not be altered in subsequent modifications to maintain the character's unique image.

[0077] In some implementations, multimodal large models are processed in parallel or sequentially to identify defect regions in multiple image regions based on a pre-set material and process knowledge base. Defect regions refer to image regions in multiple image regions that violate the process constraint rules corresponding to the material parameters.

[0078] Specifically, the material and process knowledge base is pre-stored within the system. This knowledge base contains process constraint rules for various materials, stored in a structured manner, such as a classification table or a knowledge graph. The core reason is that different materials correspond to different production processes. The same geometric feature may be produced normally in one material but may constitute a serious defect in another. For example, a right-angled corner with an interior radius of 0.8mm is prone to cracking in acrylic due to stress concentration, requiring a minimum radius of 1mm; however, in slush molding, due to its better elasticity, this angle can be demolded normally. Similarly, the minimum printable line width is 0.2mm in slush molding (usually using pad printing), but 0.8mm in plush material (usually using embroidery). Therefore, it is necessary to retrieve the corresponding process constraint rules from the knowledge base based on the material parameters input by the user to accurately identify defective areas.

[0079] In a non-limiting implementation, based on the aforementioned material and process knowledge base, the system calls a multimodal large model to perform feature measurements on each image region and generate feature data.

[0080] Specifically, for the "minimum printable linewidth" rule under slush material, in the image area of ​​the multimodal large model, such as the little finger area, the width of the line is measured by the edge detection algorithm, and the feature data obtained is 0.1mm. The system compares this feature data with the constraint range "≥0.2mm" in the process constraint rule corresponding to the material parameters.

[0081] If the feature data is outside the constraint range (0.1mm < 0.2mm), the image region to which the feature data belongs is identified as a defect region, and the deviation value between the feature data and the corresponding constraint range is recorded. For example, the deviation value here is -0.1mm, or the absolute difference of 0.1mm is recorded. If the feature data is within the constraint range, the region is not marked as a defect region.

[0082] Similarly, for the "minimum demolding angle" rule, for example, if the constraint range is ≥3 degrees, and the system measures the actual angle of a certain facade as 1.5 degrees, it is determined to be a defect area after comparison, and the deviation value is recorded as -1.5 degrees.

[0083] Step S103: Call the multimodal large model and obtain modification prompts based on the process constraint rules violated by the defect area.

[0084] In some embodiments, for example, when the material parameter is slush molding, when the system identifies a defect area, it measures the image area of ​​the little finger region: the current measured linewidth is 0.1 mm, while the constraint threshold of the process constraint rule corresponding to the slush molding material is 0.2 mm. Therefore, the image area of ​​the little finger region is detected to violate the "minimum printable linewidth" rule and is determined to be a defect area. The correction strategy for the "minimum printable linewidth" rule in the knowledge base is "morphological expansion and thickening". The system provides the location information of the defect area ("little finger region"), the violated rule ("minimum printable linewidth"), the current measurement value ("0.1 mm"), the target threshold ("0.2 mm"), and the preset correction strategy ("morphological expansion and thickening") to the multimodal large model. The multimodal large model generates modification prompts accordingly, such as: "Perform morphological expansion and thickening operation on the little finger region, increasing the linewidth from 0.1 mm to greater than 0.2 mm, while maintaining the original curvature."

[0085] In other embodiments, if the system detects that the image region of the clothing facade violates the "minimum draft angle" rule: the current measured angle is 1.5 degrees, while the constraint threshold is 3 degrees, the corresponding correction strategy in the knowledge base is "increase the draft angle". The multimodal large model then generates the modification prompt: "Increase the draft angle of the clothing facade region so that its angle with the mold opening direction is not less than 3 degrees".

[0086] Specifically, the multimodal large model generates modification instructions in natural language form based on the specific violated constraint parameters.

[0087] Step S104: Call the pre-trained image generation model, and based on the modification prompt words, redraw the defective areas belonging to the modification area to obtain the modified design drawing.

[0088] In practice, the image generation model can use the local inpainting function of a diffusion model (such as Stable Diffusion).

[0089] Specifically, the system can input the original design drawing, modification prompts, and a mask for the locked area into the image generation model. The mask for the locked area is used to protect the pixels within the locked area from modification during the redrawing process. The image generation model only performs local redrawing within the defect area covered by the modification area, based on the modification prompts, generating new pixels that conform to the process constraints, while keeping the locked area completely unchanged without any redrawing. The system outputs a modified design drawing that retains the core IP image of the trendy toy while eliminating defects that violate the process rules.

[0090] The redrawing process can be completed in one go or iteratively optimized step by step, which will be explained in detail below.

[0091] In some embodiments, for the modified design drawing, the system calculates the image feature loss gradient and manufacturability loss gradient for the defective areas belonging to the modification area based on a preset loss function. The preset loss function is a differentiable function pre-constructed by the system, based on process constraint rules in the material and process knowledge base. The system pre-designs the corresponding differentiable loss function expression according to the process constraint rules in the material and process knowledge base. For example, for the minimum demolding angle rule (constraint angle ≥ 3°), the loss function can be defined as max(0, 3° - actual angle). The loss functions corresponding to multiple rules can be combined into a total manufacturability loss function through weighted summation.

[0092] Specifically, the loss function comprises two components: visual feature loss and manufacturability loss. The visual feature loss component quantifies the visual difference between the current design and the original design within the modification area, using metrics such as perceptual loss per inch (LPIPS) or mean squared error (MSE). The manufacturability loss component quantifies the degree to which the current design violates the process constraints corresponding to the material parameters within the modification area. For example, for the "minimum printable linewidth" rule, the loss function can be defined as max(0, target linewidth - actual linewidth), meaning a positive loss occurs when the actual linewidth is less than the target linewidth; otherwise, the loss is zero.

[0093] Furthermore, the image feature loss gradient refers to the derivative of the image feature loss component in the loss function with respect to the pixel values ​​of the design drawing. It reflects the similarity between the modified design drawing and the original design drawing. Specifically, if the modified design drawing differs significantly from the original design drawing at a certain pixel in the modified area, the image feature loss gradient value at that point is large, and the gradient direction points towards the pixel change direction that can reduce the difference. The image feature loss gradient guides subsequent modifications towards a direction closer to the original design drawing. The manufacturability loss gradient, on the other hand, refers to the derivative of the manufacturability loss component in the loss function with respect to the pixel values ​​of the design drawing. It reflects the degree to which the modified design drawing satisfies the process constraints corresponding to the material parameters.

[0094] For example, if the line width of the little finger in the modified design is 0.15mm, while the process constraint requires a minimum line width of 0.2mm, then the manufacturability loss value is calculated to be 0.05 using max(0,0.2-0.15)=0.05. The system analyzes and processes each pixel within the little finger region: for pixels located at the edge of the little finger, the system determines that "if these pixels are expanded outwards, the line width will increase, and the manufacturability loss will decrease." Therefore, the manufacturability loss gradient for these pixels points to "expanding outwards," meaning the gradient direction points to the direction that can increase the line width. If the current line width already meets the constraint (e.g., 0.25mm), then the loss value is zero, and the gradient is zero. Simultaneously, the image feature loss gradient is calculated by comparing the features before and after the modification. If comparing the current little finger with the little finger in the original design, it is found that the original little finger is longer and thinner. Therefore, the image feature loss gradient points to "contracting inwards to reduce the line width." Specifically, at the pixel level, this means that edge pixels change from the line color back to the background color to restore the original state.

[0095] In some non-limiting embodiments, weighted calculation refers to the system multiplying the image feature loss gradient and the productivity loss gradient by their respective weight coefficients and then summing them to obtain the total gradient. The system uses the calculated total gradient as a guiding signal and inputs it into the image generation model. Guided by the total gradient, the image generation model adjusts the pixels in the modification area, causing the generated image to be adjusted in the direction of reducing both image feature loss and productivity loss, thus modifying the modified design and generating a new round of modified design.

[0096] The loss function L(I) includes an image feature loss component L. feat (I) and the productivity loss component L prod (I), where I is the pixel tensor of the current design, i.e., the current toy design being revised, represented in the computer as an array of pixel values. I0 represents the pixel tensor corresponding to the original design. L prodThe construction of (I) must ensure that it is differentiable with respect to any pixel of I, so that it can automatically differentiate and calculate the gradient using a large model. The productivity loss function is constructed based on process constraint rules, and its specific expression is:

[0097] L prod (I) = Σ λ k · max(0, t k – f k (I))

[0098] Among them, L prod (I) represents the manufacturability loss value corresponding to the current design drawing, used to quantify the degree to which the design drawing violates process constraints; k is the number of the process constraint rule; λ k t is the weighting coefficient for the k-th process rule, used to adjust the importance of different rules; k The target threshold required by the k-th rule, for example, a minimum line width of 0.2mm; f k This is a differentiable measurement function for the k-th rule, used to measure the corresponding characteristic values ​​from the design drawing, such as the actual linewidth. The maximum value is taken to ensure the loss value is not negative; when the design drawing meets the process constraints, the loss is 0. For the measurement method of the actual linewidth, please refer to the relevant description in step S101 above.

[0099] Taking the "minimum line width" constraint as an example: t k =0.2mm, f k (I) If the actual line width obtained by measurement is 0.1mm, then the loss is max (0, 0.2-0.1)=0.1.

[0100] Since the loss function is composed of differentiable basic operations, its partial derivative with respect to each pixel can be directly calculated by the automatic differentiation mechanism of the general deep learning framework without the need for additional custom complex solution steps, which can be directly implemented by those skilled in the art.

[0101] Furthermore, the manufacturability loss gradient is obtained by differentiating the above loss function with respect to the image pixels, and is used to represent the direction and magnitude of reducing process defects by modifying each pixel.

[0102] In practice, the above redrawing can be iterated multiple times until the preset iteration termination condition is met.

[0103] Specifically, the system recalculates the image feature loss gradient and manufacturability loss gradient for the previously modified design drawing. Since the state of the design drawing changes after each iteration (e.g., the linewidth of the defect area changes from 0.15mm to 0.25mm), the recalculated gradient values ​​also change accordingly. The image feature loss gradient is always calculated based on the difference between the current design drawing and the original design drawing, while the manufacturability loss gradient is always calculated based on the difference between the current design drawing and the target value of the process constraint rules. Furthermore, the system uses the same weighted calculation formula, but the data is based on the updated image feature loss gradient and manufacturability loss gradient obtained from the recalculation, thus obtaining the updated total gradient. The updated total gradient is input into the image generation model, which then modifies the previously modified design drawing again based on this updated total gradient, outputting a new round of modified design drawings. Each iteration is equivalent to a fine-tuning based on the original corrections.

[0104] Specifically, the image generation model used can be a text-based conditional diffusion model (e.g., StableDiffusion), which is a common existing image generation model. This embodiment of the invention does not require structural changes or retraining of the diffusion model itself. Instead, it improves the model generation process through gradient guidance, further optimizing the image based on text prompts while adhering to both image feature preservation and process constraints.

[0105] During the redrawing process, the total gradient is injected into the text conditional diffusion model during the model inference phase. The specific implementation method is as follows:

[0106] In each denoising iteration of the diffusion model, the model predicts noise based on the text encoding conditions corresponding to the modified prompt words. This application, while preserving the guiding role of the text, uses the total gradient as an additional constraint signal to correct the noise prediction results, ensuring that the image generation direction simultaneously satisfies the requirements of preserving the original IP image features and material / process constraints.

[0107] Since the total gradient is obtained by weighting the image feature loss gradient and the productivity loss gradient, it is defined in pixel space. The diffusion model's operation is based on latent space variables. Therefore, the model's built-in variational autoencoder decoder (VAE decoder) decodes the latent space variables into a pixel space image, completes the loss function calculation and gradient solution in pixel space, and uses an automatic differentiation mechanism (such as PyTorch's computation graph) to backpropagate the pixel space gradient to the latent space through the Jacobian matrix of the VAE decoder, so that the total gradient can effectively guide the model's denoising generation process.

[0108] To ensure the locked region is not modified, after each denoising iteration, the latent variables obtained by adding noise to the original design image under the same noise schedule are used to cover and replace the latent variable regions corresponding to the locked region in the diffusion model, so that the locked region image always remains consistent with the original design. Figure 1 The original text is incomplete and cannot be changed.

[0109] Specifically, in each denoising iteration t of the diffusion model, the model modifies the text encoding c of the prompt word and adjusts the current latent variable x. t Noise prediction is performed to obtain the initial predicted noise ε(x) t The total loss function L (comprising the image feature loss and the productivity loss weighted in pixel space) is applied to the current latent variable x. t The total loss function for x is calculated via backpropagation through the VAE decoder using the chain rule. t gradient x t L (x t ).

[0110] Furthermore, the initial prediction noise is corrected using the aforementioned gradient to obtain the corrected prediction noise: ε′= ε( x t , t, c) - s· x t L (x t ), where s is a preset guiding strength coefficient used to control the degree of influence of the gradient on the generated result, and its value can be set according to the correction requirements.

[0111] In practice, the system can preset a maximum number of iterations, such as 10 or 20. After each iteration, the system checks whether the current number of iterations has reached the preset maximum. If so, the iteration stops, and the last modified design is used as the final output. This condition serves as a fallback mechanism to prevent infinite loops due to other termination conditions failing to trigger, ensuring that the system outputs results within the specified number of iterations under any circumstances.

[0112] In other embodiments, during each iteration, the system calculates the manufacturability loss value of the current design drawing based on a loss function. The manufacturability loss value is a scalar, obtained by weighted summation of the loss function values ​​of various process constraint rules (such as linewidth loss, angle loss, spacing loss, etc.). The system compares this manufacturability loss value with a preset process tolerance threshold. If the manufacturability loss value is less than this threshold, the system determines that the current design drawing meets the process constraints and requires no further modification, thus terminating the iteration. This condition is the optimal termination condition, ensuring that the system stops immediately after defect repair, avoiding unnecessary computational overhead and excessive modifications.

[0113] In actual production, process parameters are usually allowed to deviate from the target value by ±5%. Therefore, the system can set the allowable deviation of each process constraint to 5% of the corresponding target value to meet the acceptable range for normal production.

[0114] Specifically, the process tolerance threshold is set individually based on the production requirements of specific process constraint rules. Different rules can use different thresholds, but the allowable deviation range is uniformly set to 5%. For example, for the minimum linewidth rule, the process tolerance threshold can be set to 0.01 mm. Based on the current design drawing, the manufacturability loss value is a scalar value calculated using the aforementioned manufacturability loss function on the current design drawing, used to quantify the overall degree to which the design drawing violates process constraints. When the manufacturability loss value is less than the set process tolerance threshold, it indicates that the current design drawing meets the mass production process requirements of the corresponding material and no further modifications are needed.

[0115] In some embodiments, the system terminates the design modification based on the maximum number of iterations and outputs a manufacturability loss value simultaneously with the modified design. The system compares this manufacturability loss value with a preset process tolerance threshold. If the manufacturability loss value is greater than the process tolerance threshold, it indicates that after the maximum number of iterations, the currently modified design still fails to fully meet the process constraint rules corresponding to the material parameters. At this time, the system sends a prompt message to the outside.

[0116] The notification message can be an internal system notification, such as a pop-up window in the user interface or a to-do item generated in the message queue; it can also be sent to a specified user or user group by calling a third-party application (such as an email service, instant messaging tool, or enterprise collaboration platform). The notification message contains the following information: the currently output modified design drawing, the current manufacturability loss value, the process tolerance threshold, and a detailed description of the unmet process constraint rules (e.g., "small finger line width 0.18mm, not meeting the minimum line width requirement of 0.2mm"). This notification message is used by humans (such as designers or process engineers) to determine whether the currently modified design drawing still has room for further modification. If the human judgment is that it can be further optimized based on the current situation (e.g., by manually adjusting or reselecting material parameters), the modified design drawing can be re-entered into the system for a new round of processing; if the human judgment is that the current design drawing can no longer meet the process requirements (e.g., the original design itself conflicts too much with the material process), the design drawing can be abandoned. Since the core value of trendy toys lies in their unique IP image, if the design still cannot meet the mass production standard after multiple automatic corrections, and if it is further modified in order to forcibly meet the process requirements, resulting in the product being far removed from the IP image, then the trendy toy loses its value.

[0117] In some implementations, the weighting coefficients can be preset values ​​used to balance the influence of the two loss gradients in the total gradient. For example, if the weighting coefficient of the image feature loss gradient is 0.7 and the weighting coefficient of the productivity loss gradient is 0.3, then the total gradient = 0.7. Image feature loss gradient +0.3 Manufacturability loss gradient. The relative magnitudes of the weighting coefficients determine whether to prioritize preserving IP features or meeting process constraints during subsequent modifications. In a typical implementation, the weighting coefficients of the image feature loss gradient are greater than those of the manufacturability loss gradient, reflecting the priority of preserving IP image features. This prevents the loss of IP recognizability due to excessive pursuit of process compliance, making it particularly suitable for trendy toy design scenarios where IP imagery is extremely sensitive and only minor modifications are permitted.

[0118] In other implementations, as the number of iterations increases, the weight coefficient of the image feature loss gradient remains constant or increases, while the weight coefficient of the manufacturability loss gradient decreases. In the early stages of iteration, process defects are more pronounced, requiring a relatively high weight for the manufacturability loss gradient to quickly eliminate them. As the number of iterations increases, process defects are gradually repaired, at which point the weight of the manufacturability loss gradient is gradually reduced to avoid excessive modification that could damage IP features. Simultaneously, the weight of the image feature loss gradient remains constant or increases accordingly to ensure that later modifications primarily maintain the image features.

[0119] In some other implementations, before performing a weighted calculation based on the updated image feature loss gradient and the updated productivity loss gradient, the system also performs the following operations: calculates the angle between the directions of the image feature loss gradient and the productivity loss gradient; if the angle is greater than a preset angle threshold, the weight coefficient of the productivity loss gradient is reduced.

[0120] Specifically, the directional angle refers to the spatial angle between the gradient vector of the image feature loss and the gradient vector of the manufacturability loss. The size of the angle reflects the degree of consistency between the two loss gradient directions: when the angle is close to 0°, the two gradient directions are the same or similar, indicating that the two objectives of "preserving IP features" and "meeting process constraints" do not conflict, and modifications can be made simultaneously along this direction; when the angle is close to 90°, the two gradient directions are orthogonal, indicating that the two objectives are independent to a certain extent; when the angle is greater than 90°, the two gradient directions are opposite, which means that there is a conflict, and the larger the angle, the more obvious the conflict. A conflict means that "preserving IP features" and "meeting process constraints" contradict each other, and modifying according to one gradient direction will increase the other loss.

[0121] In practice, the preset angle threshold is a pre-defined angle value, such as 90 degrees or 120 degrees. In each iteration, the system calculates the angle between the direction of the image feature loss gradient and the productiveity loss gradient, and compares this angle with the preset angle threshold. If the angle is greater than the preset threshold, it indicates a conflict between the two loss gradients. In this case, if the original weights are applied, the productiveity loss gradient might pull the total gradient in a direction deviating from the original design, leading to the destruction of IP features. Therefore, the system reduces the weight coefficient of the productiveity loss gradient. This reduction can be achieved by multiplying the original weight coefficient by a decay factor less than 1, setting it to a smaller preset value, or even setting it to zero.

[0122] In some implementations, the image region belonging to the modification area is further divided into at least two levels according to a preset image feature grading rule: a first-level modification area and a second-level modification area. The grading can be based on the same rules used to distinguish between the locked area and the modification area, namely, saliency detection algorithms and curvature recognition. The difference lies in the specific value of the threshold, which needs to be modified accordingly. The implementation method of the image feature grading rule will not be elaborated here. The purpose is to further differentiate the degree of association between the permitted modification areas and the core IP features.

[0123] In one implementation, based on the first threshold (e.g., 0.7) and second threshold (e.g., 15 degrees) used for locking zone determination, a lower secondary threshold is set as the boundary between the first and second levels. For example, the significance secondary threshold is set to 0.4, and the angle change secondary threshold is set to 5 degrees. Zones that do not meet the locking zone threshold but meet the secondary threshold are classified as first-level modification zones, and the rest are classified as second-level modification zones.

[0124] In other implementations, the system can also use semantic segmentation to divide the image into a first-level modification area and a second-level modification area. For example, by setting the semantic "limbs" to belong to the first level and "clothing" to belong to the second level, the system will divide the image area containing limbs into the first-level modification area and the image area containing clothing into the second-level modification area.

[0125] Similarly, it can be divided into three or more levels, with more refined hierarchical division of the modification area, in order to make more refined redrawing.

[0126] Taking the division into two levels as an example, the common division results according to the image feature classification rules are: the first level modification area corresponds to the body outline, main limb proportions, and other areas that have a relatively large impact on the overall image; the second level modification area corresponds to clothing folds, texture details, base background, and other areas that have a relatively small impact on the overall image.

[0127] Furthermore, the system assigns different weight coefficients to the image feature loss gradient for different modification levels. Specifically, the weight coefficient for the image feature loss gradient in the first-level modification area is greater than that in the second-level modification area. Therefore, during iterative modification, the system imposes a stronger constraint on maintaining image features in the first-level modification area, requiring that the modified result in this area maintain a higher similarity to the original design; while imposing a weaker constraint on maintaining image features in the second-level modification area, allowing for greater freedom in modification.

[0128] Continuing with the two-level division example, the character's arm image area is designated as the first-level modification area, and the clothing pattern area is designated as the second-level modification area. The system sets the image feature loss gradient weight coefficient for the first-level modification area to 0.8 and for the second-level modification area to 0.3. When calculating the total gradient, the system calculates the image feature loss gradient for both the first and second-level modification areas separately, multiplies each by its respective weight coefficient, and then performs a weighted sum with the manufacturability loss gradient. In this way, the arm outline is subject to strong shape-preserving constraints during modification, making it less prone to deformation; while the clothing pattern can undergo significant changes to meet process requirements.

[0129] As described above, the scheme of the first embodiment divides the locked area and the modification area by using preset hierarchical rules based on saliency detection and curvature analysis. This allows the system to identify the core features of the IP that must be protected, while redrawing defective areas in the modification area, avoiding distortion of the IP image due to excessive modification. By combining the material and process knowledge base to identify defective areas and generate modification prompts, the image generation model can specifically correct image areas that violate process constraints, thereby effectively improving the manufacturability of the design drawings while maintaining the original IP image. Because the locked area is protected, even if the modification prompts guide significant geometric adjustments during the redrawing process, it will not affect the IP identification features, ensuring the reliability of the correction results.

[0130] Figure 2 This is a schematic diagram of a manufacturability correction device for a trendy toy design drawing according to a second embodiment of the present invention. Those skilled in the art will understand that the manufacturability correction device 2 for the trendy toy design drawing in this embodiment can be used to implement the method and technical solutions described in the above embodiments.

[0131] Specifically, refer to Figure 2 The manufacturability correction device 2 for the trendy toy design drawing in this embodiment may include:

[0132] Data receiving unit 21 is used to receive the original design drawings and material parameters of the trendy toy;

[0133] The multimodal analysis unit 22 is used to call a pre-trained multimodal large model to perform region segmentation on the original design drawing and analyze the multiple image regions obtained after segmentation. The multiple image regions are divided into at least modification areas and locked areas based on the preset image feature classification rules, and defect areas in the multiple image regions are identified based on the preset material and process knowledge base. The material and process knowledge base contains process constraint rules for various materials. Defect areas refer to image regions in the multiple image regions that violate the process constraint rules corresponding to the material parameters.

[0134] The prompt word generation unit 23 is used to call the multimodal large model and obtain modification prompt words based on the process constraint rules violated by the defect area;

[0135] Image correction unit 24 is used to call a pre-trained image generation model, and based on the modification prompt words, redraw the defective areas belonging to the modification area to obtain the modified design drawing.

[0136] The process of dividing multiple image regions into at least modification and locking regions based on preset image feature grading rules includes: determining the saliency of each image region based on a preset saliency detection algorithm; identifying image regions with saliency higher than a first threshold as candidate locking regions; performing curvature analysis on multiple lines contained in the candidate locking regions, and if the rate of change of curvature of any line exceeds a second threshold, then the candidate locking region is identified as a locking region.

[0137] In specific implementation, the modules / units included in the various devices and products described in the above embodiments can be software modules / units, hardware modules / units, or a combination of both.

[0138] For example, for various devices and products applied to or integrated into a chip, each module / unit can be implemented using hardware methods such as circuits, or at least some modules / units can be implemented using software programs that run on a processor integrated within the chip, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits; for various devices and products applied to or integrated into a chip module, each module / unit can be implemented using hardware methods such as circuits, and different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components of the chip module, or at least some modules / units can be implemented using hardware methods such as circuits. The components can be implemented using software programs that run on the processor integrated within the chip module. The remaining (if any) modules / units can be implemented using hardware methods such as circuits. For various devices and products applied to or integrated into the terminal, each of its components / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or in different components within the terminal. Alternatively, at least some modules / units can be implemented using software programs that run on the processor integrated within the terminal, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits.

[0139] Furthermore, embodiments of the present invention also disclose a computer-readable storage medium, which is a non-volatile or non-transient storage medium, on which a computer program is stored, and the computer program is executed by a computer. Figure 1 The method and technical solution in the illustrated embodiment. Preferably, the storage medium may include ROM, RAM, disk, or optical disk, etc.

[0140] Furthermore, embodiments of the present invention also disclose a computer program product, including a computer program / instruction, wherein when the computer program / instruction is executed by a computer, the manufacturability modification method of any of the above-mentioned trendy toy design drawings is executed.

[0141] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for manufacturability modification of a trendy toy design, characterized in that, include: Receive original design drawings and material parameters for trendy toys; The original design drawing is segmented by calling a pre-trained multimodal large model, and the resulting image regions are analyzed. The image regions are divided into at least modification and locking regions based on a preset image feature classification rule. Defect regions are identified in the image regions based on a preset material and process knowledge base, which contains process constraint rules for various materials. The defect region refers to the image region in the image regions that violates the process constraint rules corresponding to the material parameters. The multimodal large model is invoked, and modification prompts are obtained based on the process constraint rules violated by the defect region; The pre-trained image generation model is invoked, and the defective areas belonging to the modification area are redrawn based on the modification prompt words to obtain the modified design drawing; in, The preset image feature classification rule divides the multiple image regions into at least a modification region and a locked region, including: The saliency of each image region is determined based on a pre-defined saliency detection algorithm; Image regions with a significance level higher than the first threshold are identified as candidate locking regions; Curvature analysis is performed on multiple lines contained in the candidate locking region. If the rate of change of curvature of any line exceeds a second threshold, the candidate locking region is determined as the locking region.

2. The method for manufacturability modification of trendy toy design drawings as described in claim 1, characterized in that, Also includes: For the modified design drawing, the image feature loss gradient and manufacturability loss gradient are calculated for the defect area belonging to the modification area based on the preset loss function. The image feature loss gradient reflects the similarity between the modified design drawing and the original design drawing, and the manufacturability loss gradient reflects the degree to which the modified design drawing meets the process constraint rules corresponding to the material parameters. The loss function is pre-constructed based on the process constraint rules. Calculate or obtain the weight coefficients of the image feature loss gradient and the productivity loss gradient, and perform a weighted calculation based on the image feature loss gradient and the productivity loss gradient to obtain the total gradient; The image generation model is invoked, and the modified design drawing is modified based on the total gradient.

3. The method for manufacturability modification of trendy toy design drawings as described in claim 2, characterized in that, Also includes: Repeat the following steps until the preset iteration termination condition is met: Receive the design drawing after the last modification, and update the image feature loss gradient and the productiveization loss gradient based on the preset loss function; The updated total gradient is obtained by weighting the updated image feature loss gradient with the updated productivity loss gradient. The image generation model is invoked, and the previously modified design drawing is modified based on the updated total gradient.

4. The method for manufacturability modification of trendy toy design drawings as described in claim 3, characterized in that, The preset iteration termination condition includes at least one of the following: The number of iterations has reached the preset maximum. The manufacturability loss value is less than a preset process tolerance threshold, and the manufacturability loss value is calculated based on the loss function.

5. The method for manufacturability modification of trendy toy design drawings as described in claim 3, characterized in that, When performing weighted calculations, the weight coefficients of the image feature loss gradient and the productivity loss gradient are preset fixed values, wherein the weight coefficient of the image feature loss gradient is greater than the weight coefficient of the productivity loss gradient.

6. The method for manufacturability modification of trendy toy design drawings as described in claim 3, characterized in that, As the number of iterations increases, the weight coefficient of the image feature loss gradient remains unchanged or increases, while the weight coefficient of the productivity loss gradient decreases.

7. The method for manufacturability modification of trendy toy design drawings as described in claim 3, characterized in that, Before performing the weighted calculation based on the updated image feature loss gradient and the updated productivity loss gradient, the following steps are also included: Calculate the angle between the directions of the image feature loss gradient and the productivity loss gradient; If the included angle is greater than a preset angle threshold, then the weight coefficient of the productivity loss gradient is reduced.

8. The method for manufacturability modification of trendy toy design drawings as described in claim 3, characterized in that, The modification area is divided into at least a first-level modification area and a second-level modification area according to the preset image feature classification rules. The first-level modification area and the second-level modification area are assigned different weight coefficients for the image feature loss gradient, and the weight coefficient of the image feature loss gradient of the first-level modification area is greater than the weight coefficient of the image feature loss gradient of the second-level modification area.

9. The method for manufacturability modification of trendy toy design drawings as described in claim 1, characterized in that, The method of identifying defect regions in the multiple image regions based on a pre-set material and process knowledge base includes: The multimodal large model is invoked to perform feature measurements on each image region, generating feature data; The feature data is compared with the constraint range in the process constraint rules corresponding to the material parameters; If the feature data is outside the constraint range, the image region to which the feature data belongs is identified as a defect region, and the deviation value of the feature data from the corresponding constraint range is recorded.

10. A manufacturability correction device for a trendy toy design, characterized in that, include: The data receiving unit is used to receive the original design drawings and material parameters of the trendy toy; The multimodal analysis unit is used to call a pre-trained multimodal large model to perform region segmentation on the original design drawing, and to analyze the multiple image regions obtained after segmentation. Specifically, based on a preset image feature grading rule, the multiple image regions are divided into at least a modification region and a locked region. Based on a preset material and process knowledge base, defect regions in the multiple image regions are identified. The material and process knowledge base contains process constraint rules for various materials. The defect region refers to the image region in the multiple image regions that violates the process constraint rules corresponding to the material parameters. The prompt word generation unit is used to call the multimodal large model and obtain modification prompt words based on the process constraint rules violated by the defect area; The image correction unit is used to call a pre-trained image generation model and redraw the defective areas belonging to the modification area based on the modification prompt words to obtain the modified design drawing. The step of dividing the multiple image regions into at least modification regions and lock regions based on preset image feature grading rules includes: determining the saliency of each image region based on a preset saliency detection algorithm; identifying image regions with saliency higher than a first threshold as candidate lock regions; performing curvature analysis on multiple lines contained in the candidate lock regions, and if the rate of change of curvature of any line exceeds a second threshold, then the candidate lock region is identified as the lock region.

11. A computer-readable storage medium, said computer-readable storage medium being a non-volatile storage medium or a non-transient storage medium, having stored thereon a computer program, characterized in that, The computer program is executed by a computer to perform the manufacturability modification method of the trendy toy design drawing according to any one of claims 1 to 9.

12. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the computer, the manufacturability modification method of the trendy toy design drawing according to any one of claims 1 to 9 is executed.

Images