Drawing method and system based on AI model

Abstract

The application discloses a kind of based on AI model's drawing method and system, method includes: multiple modal drawing demand data is converted into structured drawing demand data;Rule matching is carried out, and professional feature rule set is generated;Filter parameter adjustment strategy is dynamically quantified assignment to core inference parameter, and drawing core inference parameter set is generated;According to drawing core inference parameter set, generate layered drawing basic layer set;To layered drawing basic layer set is spliced processing, and the neural rendering of fusion hand-drawing brush touch, generate initial edition complete drawing work;Local parameter adjustment is carried out, and the final drawing work and corresponding editable drawing parameter file are generated.The application is converted into structured data by multiple modal drawing demand, and the whole process design of model parameter dynamic assignment, layered layer generation, brush touch fusion rendering and incremental optimization is carried out after art rule matching, improves the demand fit degree and reliability of AI drawing.

Description

Technical Field

[0001] This application relates to the field of AI drawing technology, and in particular to a drawing method and system based on an AI model. Background Technology

[0002] In the interdisciplinary field of computer vision, digital graphics, and deep learning, AI-based drawing is an important research and application area. Currently, the integration of multimodal requirements analysis and the technical systems of fine arts within this field still needs further development, and there is still room for exploration in the technical collaboration of layered generation, style transfer, and hand-drawn texture rendering. Meanwhile, various industries are continuously increasing their demands for the professionalism, personalization, and editability of AI-based drawing. The existing technology's integration and adaptation capabilities are insufficient to meet the cross-domain application needs, and breakthroughs are urgently needed in the professionalism and flexibility of technology implementation. Therefore, research on AI model-based drawing technologies is necessary.

[0003] In the prior art, Chinese patent CN119151793A discloses a method and system for generating descriptions of controllable, high-quality AI-generated drawings. The method includes using natural language processing technology to analyze and identify the user's drawing intentions and extract key drawing elements; understanding the meaning of the elements and the relationships between them based on a deep learning neural network model, and expanding and supplementing the user's drawing intentions based on a generative adversarial network to generate drawing prompts in various directions; generating images that conform to the descriptions of the drawing prompts using a diffusion model image generation algorithm based on the drawing prompts, and obtaining images of various styles that meet the expected quality standards by introducing an image quality assessment system.

[0004] However, the aforementioned existing technologies only focus on generating drawing prompts through natural language processing, generating images using diffusion models, and quality assessment. They do not incorporate professional art rules for precise constraints, lack a systematic approach to multimodal demand analysis, and have insufficient flexibility in parameter configuration. As a result, they are difficult to accurately match users' actual drawing needs, and the demand fit and reliability of AI drawing need to be improved. Summary of the Invention

[0005] This application provides a drawing method and system based on an AI model to address the problem that the suitability and reliability of AI drawing in the prior art need to be improved.

[0006] On the one hand, this application provides a drawing method based on an AI model, including the following steps: Step 1: Obtain multimodal drawing requirement data and use a preset art domain ontology library to transform the multimodal drawing requirement data into structured drawing requirement data.

[0007] Step two: Use a feature similarity matching algorithm to match the structured drawing requirement data with various rules in the preset art professional rule library to generate a professional feature rule set.

[0008] Step 3: Based on the professional feature rule set, select the corresponding parameter adjustment strategy from the preset style-parameter adaptation strategy library to dynamically quantify and assign values ​​to the core inference parameters of the AI ​​drawing model, thereby generating the drawing core inference parameter set of the AI ​​drawing model.

[0009] Step four: The AI ​​drawing model divides the model into independent layers according to the professional feature rule set, and generates the base images of each independent layer in sequence according to the drawing core inference parameter set, thus obtaining a layered drawing base layer set.

[0010] Step 5: Use a seamless layer fusion algorithm to stitch together the layered drawing base layer set, and then perform neural rendering of the stitched drawing base image with hand-drawn brushstrokes to generate the first version of the complete drawing artwork.

[0011] Step six: Use an incremental optimization algorithm to adjust the local parameters of the initial complete drawing to generate the final drawing and the corresponding editable drawing parameter file.

[0012] In one possible implementation, in step one, the multimodal drawing requirement data includes text-based drawing requirement data and visual-based drawing requirement data.

[0013] Semantic feature extraction, keyword clustering, and sentiment feature extraction are performed on the text-based drawing requirement data to obtain a text-based requirement feature vector containing a semantic feature vector, a set of core requirement keywords, and sentiment feature values.

[0014] The visual drawing requirement data is subjected to feature point extraction, contour feature clustering, color histogram extraction, and visual style feature extraction to obtain a visual requirement feature vector containing visual feature point set, contour feature clustering result, color histogram, and visual style feature.

[0015] A preset art domain ontology library is loaded. The art domain ontology library contains several core art dimensions. Each core art dimension has a standard feature vector library. The cosine similarity algorithm is used to match the text-type requirement feature vector and the visual-type requirement feature vector with the standard feature vector in the standard feature vector library one by one. The standard feature vector with the highest matching degree is selected for each core art dimension and transformed into the corresponding structured drawing requirement data.

[0016] In one possible implementation, in step two, the dimensional rule set of the art professional rule base corresponds to the core art dimension of the art domain ontology in step one.

[0017] The feature similarity matching algorithm is used to match the structured drawing requirements data of each core art dimension with the various rules of the corresponding dimension rule set in the preset art professional rule library, and to select a preliminary art rule set.

[0018] A lightweight embedding algorithm is used to transform the preliminary art rule set into rule parameters that can be recognized by the AI ​​drawing model, and an association index between the rule parameters and the inference layer of the AI ​​drawing model is established to generate a professional feature rule set with a mapping relationship.

[0019] In one possible implementation, in step three, core style data is extracted from the set of professional feature rules, including art style type, requirements for generating image details, requirements for fusion of multiple style features, and requirements for light and shadow layers.

[0020] Based on the art style type, retrieve the corresponding style parameter adjustment strategy from the preset style-parameter adaptation strategy library.

[0021] Based on the parameter adjustment strategy, and in conjunction with the requirements for generating image details, the requirements for fusion of multiple style features, and the requirements for light and shadow layers, the core inference parameters of the AI ​​drawing model are dynamically quantified and assigned values ​​to generate the drawing core inference parameter set of the AI ​​drawing model.

[0022] The core inference parameters include generation step size, feature fusion weight, style transfer strength, and detail generation threshold.

[0023] In one possible implementation, in step four, visual hierarchy rules are extracted from the set of professional feature rules, and the AI ​​drawing model divides the model into several independent layers according to the visual hierarchy rules.

[0024] The AI ​​drawing model's inference layer generates the base images of each independent layer sequentially based on the core drawing inference parameter set, thus obtaining a layered drawing base layer set.

[0025] In one possible implementation, in step four, after obtaining the hierarchical drawing base layer set, a feature transfer algorithm is used to perform multi-style fusion processing on the hierarchical drawing base layer set.

[0026] In one possible implementation, step five, the neural rendering that integrates hand-drawn brushstrokes, includes: Brushstroke feature rules are extracted from the professional feature rule set, and suitable hand-drawn brushstroke features are selected from the preset hand-drawn brushstroke feature library according to the brushstroke feature rules.

[0027] By integrating the characteristics of hand-drawn brushstrokes into the neural rendering algorithm, the basic drawing image obtained from the splicing process is rendered to generate the first version of the complete drawing.

[0028] In one possible implementation, step six involves constructing a multi-dimensional aesthetic evaluation system, including the rationality of color matching, the standardization of composition, the accuracy of perspective, the visual balance, and the stylistic fit.

[0029] The initial complete drawing is scored based on the aforementioned multi-dimensional aesthetic evaluation system. If the score of any dimension in the multi-dimensional aesthetic evaluation system is lower than the preset qualified threshold, a corresponding local parameter adjustment requirement is generated.

[0030] An incremental optimization algorithm is used to adjust the local parameters of the initial complete drawing based on the local parameter adjustment requirements, generating the final drawing and the corresponding editable drawing parameter file.

[0031] On the other hand, this application provides a drawing system based on an AI model, which adopts the above-mentioned drawing method based on an AI model, including: a drawing requirement analysis module, a feature rule mapping module, a model parameter generation module, a layered layer generation module, a drawing screen rendering module, and a drawing work optimization module.

[0032] The drawing requirement parsing module is configured to: acquire multimodal drawing requirement data and convert the multimodal drawing requirement data into structured drawing requirement data using a preset art domain ontology library.

[0033] The feature rule mapping module is configured to use a feature similarity matching algorithm to match the structured drawing requirement data with various rules in a preset art professional rule library to generate a professional feature rule set.

[0034] The model parameter generation module is configured to: based on the professional feature rule set, select corresponding parameter adjustment strategies from the preset style-parameter adaptation strategy library to dynamically quantify and assign values ​​to the core inference parameters of the AI ​​drawing model, thereby generating the drawing core inference parameter set of the AI ​​drawing model.

[0035] The layer generation module is configured such that the AI ​​drawing model divides independent layers according to the professional feature rule set, and generates the base images of each independent layer in sequence according to the drawing core inference parameter set, thereby obtaining a layered drawing base layer set.

[0036] The drawing rendering module is configured to: use a seamless layer fusion algorithm to stitch together the layered drawing base layer set, and then perform neural rendering of the stitched drawing base image with hand-drawn brushstrokes to generate a first-version complete drawing artwork.

[0037] The drawing optimization module is configured to use an incremental optimization algorithm to adjust local parameters of the initial complete drawing, generating the final drawing and the corresponding editable drawing parameter file.

[0038] The AI-based drawing method and system described in this application have the following advantages: By transforming multimodal drawing requirements into structured data, and through a full-process design including art rule matching, dynamic assignment of model parameters, generation of layered layers, brush stroke fusion rendering, and incremental optimization, the AI ​​drawing's ability to meet the needs and its reliability have been improved.

[0039] By extracting unique feature vectors from multimodal demand data in text and visual categories, and combining them with standard feature vectors from an ontology library in the art field, cosine similarity is used for precise matching to transform the data into structured data. This improves the professionalism and accuracy of demand data analysis and lays a reliable foundation for subsequent art rule matching and model parameter adaptation.

[0040] By aligning the dimensions of the art professional rule base with the core dimensions of the art domain ontology, filtering rules through feature similarity matching, and then converting them into AI-recognizable rule parameters using a lightweight embedding algorithm and establishing an inference layer association index, the adaptability of art rules to AI drawing models is improved, while also enhancing the model's efficiency in calling art rules and its recognition accuracy.

[0041] By extracting four types of core style data from the professional feature rule set, retrieving the corresponding parameter adjustment strategies, and dynamically quantifying and assigning values ​​to the four core inference parameters in combination with the requirements of detail, fusion, and lighting, personalized adaptation of model parameters is achieved. This improves the matching accuracy between AI drawing model parameters and users' personalized drawing needs, making parameter configuration more in line with actual drawing requirements.

[0042] By extracting visual hierarchy rules from a set of professional feature rules to divide the layers into independent layers, the AI ​​drawing model's inference layer generates the basic images of each layer sequentially based on the core drawing inference parameter set. This makes the visual hierarchy division of the drawing conform to professional art standards, improving the visual hierarchy expression of AI drawing works and the professionalism and orderliness of layer generation.

[0043] After generating a basic set of layered drawing layers, a feature transfer algorithm is used to perform multi-style fusion processing on each independent layer. This achieves accurate integration and fusion of different art styles on each layer, avoiding style conflicts and abrupt fusion boundaries, and improving the naturalness of style fusion and overall visual unity of multi-style drawing works.

[0044] By extracting brushstroke feature rules from a professional feature rule set, selecting suitable features from a hand-drawn brushstroke feature library, and integrating them into a neural rendering algorithm, the basic drawing image is rendered. This allows AI-generated drawings to combine the accuracy of algorithm generation with the artistic quality of hand-drawn art, improving the artistic expression and visual richness of the drawings and meeting the aesthetic requirements of fine art creation.

[0045] By constructing a five-dimensional aesthetic evaluation system to score the initial version of the artwork, local adjustment requests are generated for dimensions that are below the threshold. An incremental optimization algorithm is used to make precise local adjustments, avoiding the need to redraw the entire image. This improves the artistic professionalism of the final artwork and outputs an editable parameter file, thereby improving the efficiency of subsequent secondary optimization of the artwork. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating a drawing method based on an AI model, as provided in an embodiment of this application. Detailed Implementation

[0048] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] like Figure 1 As shown in the figure, this application provides a drawing method based on an AI model, including the following steps: Step 1: Obtain multimodal drawing requirement data and use a preset art domain ontology library to transform the multimodal drawing requirement data into structured drawing requirement data.

[0050] Step two: Use a feature similarity matching algorithm to match the structured drawing requirement data with various rules in the preset art professional rule library to generate a professional feature rule set.

[0051] Step 3: Based on the professional feature rule set, select the corresponding parameter adjustment strategy from the preset style-parameter adaptation strategy library to dynamically quantify and assign values ​​to the core inference parameters of the AI ​​drawing model, thereby generating the drawing core inference parameter set of the AI ​​drawing model.

[0052] Step four: The AI ​​drawing model divides the model into independent layers according to the professional feature rule set, and generates the base images of each independent layer in sequence according to the drawing core inference parameter set, thus obtaining a layered drawing base layer set.

[0053] Step 5: Use a seamless layer fusion algorithm to stitch together the layered drawing base layer set, and then perform neural rendering of the stitched drawing base image with hand-drawn brushstrokes to generate the first version of the complete drawing artwork.

[0054] Step six: Use an incremental optimization algorithm to adjust the local parameters of the initial complete drawing to generate the final drawing and the corresponding editable drawing parameter file.

[0055] For example, in step one, the multimodal drawing requirement data includes text-based drawing requirement data and visual drawing requirement data.

[0056] Semantic feature extraction, keyword clustering, and sentiment feature extraction are performed on the text-based drawing requirement data to obtain a text-based requirement feature vector containing a semantic feature vector, a set of core requirement keywords, and sentiment feature values.

[0057] The visual drawing requirement data is subjected to feature point extraction, contour feature clustering, color histogram extraction, and visual style feature extraction to obtain a visual requirement feature vector containing visual feature point set, contour feature clustering result, color histogram, and visual style feature.

[0058] A preset art domain ontology library is loaded. The art domain ontology library contains several core art dimensions. Each core art dimension has a standard feature vector library. The cosine similarity algorithm is used to match the text-type requirement feature vector and the visual-type requirement feature vector with the standard feature vector in the standard feature vector library one by one. The standard feature vector with the highest matching degree is selected for each core art dimension and transformed into the corresponding structured drawing requirement data.

[0059] Specifically, in this embodiment, the text-based drawing requirement data is preprocessed, including removing meaningless interjections / punctuation marks, performing unified word segmentation between Chinese and English, and converting between simplified and traditional Chinese characters.

[0060] Preprocess the visual drawing requirements data in advance, including unifying image resolution, unifying format, and removing image watermarks / noise.

[0061] The BERT pre-trained language model is used to extract semantic features from text-based drawing requirement data to obtain semantic feature vectors, which are used to represent the core semantics of user needs (e.g., the core semantics of "warm autumn forest" is "autumn + forest + warm tone"). The K-means clustering algorithm is used to cluster keywords in the text-based drawing requirement data to extract a set of core requirement keywords (keyword weights are labeled; in this embodiment, keywords with a weight ≥ 0.5 are set as core keywords). The TextCNN sentiment model is used to extract sentiment features from the text-based drawing requirement data to obtain sentiment feature values ​​(in this embodiment, they are divided into positive, neutral, and negative, where positive corresponds to a sentiment intensity of 0.6-1, neutral corresponds to a sentiment intensity of 0.4-0.6, and negative corresponds to a sentiment intensity of 0-0.4, such as "healing sunset", which has a sentiment intensity of 0.8, indicating a positive value).

[0062] The SIFT scale-invariant feature transform algorithm is used to extract feature points from the visual drawing requirement data, resulting in a visual feature point set (labeled with the coordinates and number of feature points), which is used to represent the core visual information of the image, such as the contour and main elements. The Canny edge detection algorithm and K-means clustering algorithm are used to cluster the contour features of the visual drawing requirement data. The Canny edge detection algorithm is used to extract image contour features, and the K-means clustering algorithm is used to cluster the image contour features. The resulting contour feature clustering results are used to represent the compositional skeleton of the image. A color histogram is extracted from the visual drawing requirement data based on the HSV color space to represent the main color tone and color distribution ratio of the image. A lightweight image style recognition model is used to extract visual style features from the visual drawing requirement data from four basic dimensions: brushstroke, tone, composition, and texture, resulting in visual style features (such as realistic style, traditional Chinese style, and cartoon style).

[0063] The feature extraction algorithms and models mentioned above all use existing algorithms and models, which will not be elaborated on here.

[0064] The semantic feature vector, the set of core requirement keywords, and the sentiment feature value are processed in a unified dimension to form a text-based requirement feature vector; the visual feature point set, the contour feature clustering result, the color histogram, and the visual style feature are processed in a unified dimension to form a visual-based requirement feature vector.

[0065] In this embodiment, the art domain ontology library has six core art dimensions, including composition type, color scheme, brushstroke style, subject of the image, visual hierarchy, and perspective rules. Each core art dimension has a standard feature vector library, and each standard feature vector library includes several pairs of standard feature vectors, corresponding to several sub-dimensions of the core art dimension. For example, the color scheme has three sub-dimensions: warm color, cool color, and neutral color. Each sub-dimension has a pair of standard feature vectors, which are consistent with the dimensions of the text-based requirement feature vectors and the visual-based requirement feature vectors, respectively. Each sub-dimension also has a corresponding quantization parameter range.

[0066] Using the cosine similarity algorithm, the text-type requirement feature vector and the visual-type requirement feature vector are matched one by one with the standard feature vectors in the standard feature vector library. For each core art dimension, the standard feature vector with the highest matching degree is selected, that is, a sub-dimension of each core art dimension is matched, and the corresponding quantitative parameter range is obtained as the structured drawing requirement data of that core art dimension.

[0067] For example, in step two, the dimensional rule set of the art professional rule base corresponds to the core art dimensions of the art domain ontology in step one.

[0068] The feature similarity matching algorithm is used to match the structured drawing requirements data of each core art dimension with the various rules of the corresponding dimension rule set in the preset art professional rule library, and to select a preliminary art rule set.

[0069] A lightweight embedding algorithm is used to transform the preliminary art rule set into rule parameters that can be recognized by the AI ​​drawing model, and an association index between the rule parameters and the inference layer of the AI ​​drawing model is established to generate a professional feature rule set with a mapping relationship.

[0070] Specifically, in this embodiment, the art professional rule base has six categories of dimensional rule sets, including the composition professional rule set (corresponding to composition types, including a complete set of professional rules for composition subtypes such as central composition / rule of thirds / golden ratio), the color matching rule set (corresponding to color system, including a complete set of professional rules for color subtypes such as warm / cool / neutral color), the brushstroke matching rule set (corresponding to brushstroke style, including a complete set of professional rules for brushstroke subtypes such as thick paint / ink wash / hatching / smudging), the visual balance rule set (corresponding to the main subject of the picture, including a complete set of professional rules for the main subject subtypes such as nature / figure / still life), the visual hierarchy rule set (corresponding to visual hierarchy, including several visual layers, each visual layer including corresponding layer sub-rules and inter-layer transition rules), and the perspective professional rule set (corresponding to perspective rules, including a complete set of professional rules for perspective subtypes such as one point / two points / no perspective).

[0071] In this embodiment, the structured drawing requirements data of each core art dimension are converted into quantitative parameter feature vectors. A feature similarity matching algorithm (in this embodiment, Euclidean distance is used to calculate the feature matching degree) is used to match the quantitative parameter feature vectors of each core art dimension with the rule feature vectors of various rules in the corresponding dimension rule set in the preset art professional rule library. The rules corresponding to each core art dimension are then selected and integrated to obtain a preliminary art rule set.

[0072] In this embodiment, the lightweight embedding algorithm includes structured encoding and low-dimensional embedding of the preliminary art rule set. Specifically, the qualitative subtypes of the preliminary art rule set are mapped to discrete numerical codes, and the quantitative parameter intervals are normalized to continuous values ​​of 0-1 to form a high-dimensional rule feature matrix. Then, the high-dimensional rule feature matrix is ​​compressed into a low-dimensional dense vector that can be recognized by the AI ​​drawing model (i.e., rule parameters that can be recognized by the AI ​​drawing model), and an association index between the rule parameters and the inference layer of the AI ​​drawing model is established to generate a professional feature rule set with a mapping relationship.

[0073] For example, in step three, core style data is extracted from the set of professional feature rules, including art style type, requirements for generating image details, requirements for fusion of multiple style features, and requirements for light and shadow layers.

[0074] Based on the art style type, retrieve the corresponding style parameter adjustment strategy from the preset style-parameter adaptation strategy library.

[0075] Based on the parameter adjustment strategy, and in conjunction with the requirements for generating image details, the requirements for fusion of multiple style features, and the requirements for light and shadow layers, the core inference parameters of the AI ​​drawing model are dynamically quantified and assigned values ​​to generate the drawing core inference parameter set of the AI ​​drawing model.

[0076] The core inference parameters include generation step size, feature fusion weight, style transfer strength, and detail generation threshold.

[0077] Specifically, in this embodiment, the process of extracting art style types is as follows: traverse all sub-rule groups of the professional feature rule set, extract all qualitative labels with the [style feature] extraction identifier, remove duplicate labels, and obtain the art style type (labeling the number of styles, style name, and style priority).

[0078] The extraction process for the image detail generation requirements is as follows: Traverse the professional feature rule set from the dimensions of brushstroke style, visual hierarchy, and image subject, extract all sub-rules with extraction tags of [detail level / detail density / detail layer parameters] qualitative descriptions and quantitative parameter ranges, integrate the sub-rules qualitative descriptions and quantitative parameter ranges of the three dimensions into unified image detail generation requirements. If there is a conflict among the three dimensions, the brushstroke style dimension shall prevail, followed by the visual hierarchy dimension, and then the image subject dimension.

[0079] The extraction process for multi-style feature fusion requirements is as follows: Traverse the professional feature rule set from the visual hierarchy dimension and the color system dimension to extract the qualitative requirements for style fusion, the layer on which the core features of each style are mounted, and the style feature compatibility threshold, thus forming multi-style feature fusion requirements.

[0080] The extraction process of the light and shadow hierarchy requirements is as follows: directly extract all the contents of the light and shadow hierarchy sub-rule group from the visual hierarchy dimension of the professional feature rule set to obtain the light and shadow hierarchy requirements.

[0081] In this embodiment, the style-parameter adaptation strategy library stores a one-to-one correspondence between art style type and parameter adjustment strategy. Based on the art style type, the corresponding style parameter adjustment strategy is retrieved from the style-parameter adaptation strategy library, including the basic adjustment range and basic weights for four core inference parameters: generation step size, feature fusion weight, style transfer strength, and detail generation threshold.

[0082] The basic weights of feature fusion are adjusted according to the requirements of multi-style feature fusion, including: when the requirement of multi-style feature fusion is multi-style, the basic weight ratio of feature fusion is adjusted according to the proportion of the primary and secondary styles; when the requirement of multi-style feature fusion is single-style, the basic weight of feature fusion is locked.

[0083] The generation step size, style transfer intensity, and detail generation threshold are adjusted sequentially within the basic adjustment range based on the requirements for image detail generation and lighting level. The adjustment logic is as follows: the higher the requirements for image detail generation and lighting level, the closer the values ​​of generation step size, style transfer intensity, and detail generation threshold are to the basic adjustment range.

[0084] The unique quantized values ​​of the four adjusted core inference parameters are integrated to generate the core inference parameter set for the AI ​​drawing model.

[0085] For example, in step four, visual hierarchy rules are extracted from the set of professional feature rules, and the AI ​​drawing model divides the model into several independent layers according to the visual hierarchy rules.

[0086] The AI ​​drawing model's inference layer generates the base images of each independent layer sequentially based on the core drawing inference parameter set, thus obtaining a layered drawing base layer set.

[0087] Specifically, in this embodiment, the visual hierarchy rules include the definition of the number of layers (in this embodiment, the range of visual layers is set to 1 to 5), the hierarchical relationship between layers (the background layer is the bottom layer, the main body layer is superimposed on the background layer, the detail layer is superimposed on the main body layer, and the light and shadow layer is the top layer covering all layers), the transition rules between layers (the pixel blending parameter range of the edges of each layer and the transition width requirements), and the layered pixel range (the canvas pixel ratio and coordinate range requirements of each layer).

[0088] AI drawing models divide a number of independent layers into several layers based on the definition of the number of layers, the hierarchical relationship between layers, the transition rules between layers, and the pixel range of each layer in the visual hierarchy rules (for example, when there are 4 layers, the 4 independent layers are: background layer - main body layer - detail layer - light and shadow layer).

[0089] Load the core inference parameter set for drawing, assign a unique parameter combination to each independent layer, and generate the basic images of each independent layer from the bottom to the top layer in the inference layer of the AI ​​drawing model according to the inter-layer hierarchy relationship in the visual hierarchy rules, thus obtaining the layered drawing basic layer set.

[0090] For example, in step four, after obtaining the layered drawing base layer set, a feature transfer algorithm is used to perform multi-style fusion processing on the layered drawing base layer set.

[0091] Specifically, in this embodiment, when the multi-style feature fusion requirement is a single style, no multi-style fusion processing is performed. When the multi-style feature fusion requirement is multiple styles, the core artistic features of different styles are extracted using existing feature transfer algorithms, and the features are transferred to the corresponding layers (such as the main body layer of the Chinese style + the background layer of the oil painting style). At the same time, feature smoothing processing is performed at the edge fusion of the layers to eliminate style boundaries and achieve natural fusion of multiple styles.

[0092] For example, in step five, the neural rendering that integrates hand-drawn brushstrokes includes: Brushstroke feature rules are extracted from the professional feature rule set, and suitable hand-drawn brushstroke features are selected from the preset hand-drawn brushstroke feature library according to the brushstroke feature rules.

[0093] By integrating the characteristics of hand-drawn brushstrokes into the neural rendering algorithm, the basic drawing image obtained from the splicing process is rendered to generate the first version of the complete drawing.

[0094] Specifically, in this embodiment, in step five, the basic layer set of layered drawing is first retrieved, and existing seamless layer fusion algorithms (such as Poisson fusion, Laplacian fusion, feathering fusion, etc., which can be selected according to the actual situation) are used to stitch together the base images of each independent layer in the basic layer set of layered drawing. During the stitching process, the edge pixels of the layers are subjected to gradient transition processing to eliminate layer stitching traces and ensure the overall visual continuity of the image.

[0095] In this embodiment, the hand-drawn brushstroke feature library contains mainstream hand-drawn brushstroke features in the art field (oil painting impasto brushstrokes, traditional Chinese ink painting brushstrokes, sketch hatching brushstrokes, watercolor shading brushstrokes, cartoon flat painting brushstrokes, etc.). Each hand-drawn brushstroke feature contains quantifiable parameters such as texture, density, intensity, and direction. Brushstroke feature rules are extracted from a professional feature rule set, and suitable hand-drawn brushstroke features are selected from the hand-drawn brushstroke feature library according to these rules.

[0096] The adapted hand-drawn brushstroke features are embedded into the rendering kernel parameter nodes of the neural rendering algorithm (the neural rendering algorithm is selected according to the actual situation, including NeRF-based neural rendering, GAN-based generative rendering, CNN-based texture rendering, etc.) to achieve deep integration of brushstroke features and rendering algorithms; for the basic drawing image obtained by splicing, the corresponding brushstroke rendering parameters are bound to the layers to ensure that the brushstroke performance of different layers conforms to the rules of art.

[0097] Based on a neural rendering algorithm that integrates brushstroke features, the basic drawing image is rendered layer by layer in a visual hierarchy from bottom to top. After rendering, the entire image is uniformly color-calibrated to ensure the visual consistency of brushstrokes, tones, and lighting across layers, eliminating the sense of boundary between layers and obtaining a first-version complete drawing.

[0098] For example, in step six, a multi-dimensional aesthetic evaluation system is constructed, including the rationality of color matching, the standardization of composition, the accuracy of perspective, the visual balance, and the style fit.

[0099] The initial complete drawing is scored based on the aforementioned multi-dimensional aesthetic evaluation system. If the score of any dimension in the multi-dimensional aesthetic evaluation system is lower than the preset qualified threshold, a corresponding local parameter adjustment requirement is generated.

[0100] An incremental optimization algorithm is used to adjust the local parameters of the initial complete drawing based on the local parameter adjustment requirements, generating the final drawing and the corresponding editable drawing parameter file.

[0101] Specifically, in this embodiment, the scoring range for each dimension of the multi-dimensional evaluation system is uniformly set to 0-10 points, and the preset passing threshold is set to 7 points; at the same time, a corresponding quantitative evaluation index system is built for each dimension to realize the transformation from qualitative evaluation to quantitative scoring: The rationality of color matching is based on the HSV color space, and is comprehensively scored by calculating quantitative indicators such as the proportion of the main color tone in the picture, the ratio of complementary / adjacent colors, and the uniformity of color brightness / saturation.

[0102] The composition is standardized by comparing it with the composition rules in the art professional rules collection. The composition ratio of the elements, the coordinate positioning of the main elements, the proportion of white space, and other indicators are checked to see if they meet the requirements of the rules and are scored quantitatively.

[0103] Perspective accuracy is assessed based on professional perspective rules, and is evaluated on indicators such as the number and position of vanishing points, the relationship of line perspective extension, and the proportional adaptation of near and far objects.

[0104] Visual balance combines a set of visual balance rules to evaluate and score indicators such as the visual proportion of the subject and background, the distribution of the center of gravity of elements, and the visual attractiveness of the visual focus.

[0105] Style fit is calculated by comparing the visual style features, brushstroke features, and color features of the initial artwork with the features of the extracted art style type, and then quantifying and scoring the similarity scores.

[0106] The initial complete drawing is scored based on a multi-dimensional aesthetic evaluation system. If the score of any dimension in the multi-dimensional aesthetic evaluation system is lower than the preset qualified threshold (7 points), a corresponding local parameter adjustment requirement is generated. The local parameter adjustment requirement includes the name of the dimension to be optimized, the current actual score, the target optimization score, the associated core inference parameters of the AI ​​drawing model, and the adjustable range of the parameters.

[0107] To address the need for local parameter adjustments, the incremental optimization algorithm employs a combination of locking the target parameters and gradient incremental iteration. The specific process is as follows: First, the local parameter adjustment requirements are mapped to the adjustable parameter set of the AI ​​drawing model. The core inference parameters corresponding to all target dimensions are kept unchanged, and only the core inference parameters associated with non-target dimensions are extracted as parameters to be optimized (e.g., feature fusion weights and color histogram quantization parameters related to the color scheme for non-target color matching; generation step size and layered pixel interval parameters related to the composition type for non-target composition). Second, gradient descent is used to optimize the parameters. The parameters are adjusted incrementally in small steps. In this embodiment, the step size of each parameter adjustment is set to 5% of the basic adjustment range. After each adjustment, only the local area of ​​the image corresponding to the dimension that does not meet the standard is re-rendered and the quantitative score of that dimension is redone. When the score of that dimension reaches the qualified threshold of 7 points, or the number of parameter iteration adjustments reaches the preset upper limit of 5 times, the parameter optimization of that dimension is stopped. If there are multiple dimensions that do not meet the standard, incremental optimization is carried out in the order of style fit > composition standardization > perspective accuracy > visual balance > color matching rationality to avoid image parameter conflicts caused by simultaneous adjustment of multiple dimensions.

[0108] After optimizing all parameters in the unmet dimensions, the final drawing and its corresponding editable drawing parameter file are generated. The editable drawing parameter file stores the core parameter information of the entire process in a standardized key-value pair format. Specifically, it includes structured drawing requirement data, professional feature rule set, initial drawing core inference parameter set, exclusive parameter combination for each independent layer, quantitative parameters such as texture / density / strength / direction of hand-drawn brushstroke features, and the final parameter values ​​after incremental optimization. The file supports individual retrieval and modification of parameters, enabling secondary editing, local fine-tuning, and style redrawing of the final drawing.

[0109] This application also provides an AI model-based drawing system, which employs the above-mentioned AI model-based drawing method, including: a drawing requirement analysis module, a feature rule mapping module, a model parameter generation module, a layered layer generation module, a drawing screen rendering module, and a drawing work optimization module.

[0110] The drawing requirement parsing module is configured to: acquire multimodal drawing requirement data and convert the multimodal drawing requirement data into structured drawing requirement data using a preset art domain ontology library.

[0111] The feature rule mapping module is configured to use a feature similarity matching algorithm to match the structured drawing requirement data with various rules in a preset art professional rule library to generate a professional feature rule set.

[0112] The model parameter generation module is configured to: based on the professional feature rule set, select corresponding parameter adjustment strategies from the preset style-parameter adaptation strategy library to dynamically quantify and assign values ​​to the core inference parameters of the AI ​​drawing model, thereby generating the drawing core inference parameter set of the AI ​​drawing model.

[0113] The layer generation module is configured such that the AI ​​drawing model divides independent layers according to the professional feature rule set, and generates the base images of each independent layer in sequence according to the drawing core inference parameter set, thereby obtaining a layered drawing base layer set.

[0114] The drawing rendering module is configured to: use a seamless layer fusion algorithm to stitch together the layered drawing base layer set, and then perform neural rendering of the stitched drawing base image with hand-drawn brushstrokes to generate a first-version complete drawing artwork.

[0115] The drawing optimization module is configured to use an incremental optimization algorithm to adjust local parameters of the initial complete drawing, generating the final drawing and the corresponding editable drawing parameter file.

[0116] This application's embodiments improve the suitability and reliability of AI drawing by transforming multimodal drawing requirements into structured data and then implementing a full-process design that includes art rule matching, dynamic assignment of model parameters, generation of layered layers, brush stroke fusion rendering, and incremental optimization.

[0117] By extracting unique feature vectors from multimodal demand data in text and visual categories, and combining them with standard feature vectors from an ontology library in the art field, cosine similarity is used for precise matching to transform the data into structured data. This improves the professionalism and accuracy of demand data analysis and lays a reliable foundation for subsequent art rule matching and model parameter adaptation.

[0118] By aligning the dimensions of the art professional rule base with the core dimensions of the art domain ontology, filtering rules through feature similarity matching, and then converting them into AI-recognizable rule parameters using a lightweight embedding algorithm and establishing an inference layer association index, the adaptability of art rules to AI drawing models is improved, while also enhancing the model's efficiency in calling art rules and its recognition accuracy.

[0119] By extracting four types of core style data from the professional feature rule set, retrieving the corresponding parameter adjustment strategies, and dynamically quantifying and assigning values ​​to the four core inference parameters in combination with the requirements of detail, fusion, and lighting, personalized adaptation of model parameters is achieved. This improves the matching accuracy between AI drawing model parameters and users' personalized drawing needs, making parameter configuration more in line with actual drawing requirements.

[0120] By extracting visual hierarchy rules from a set of professional feature rules to divide the layers into independent layers, the AI ​​drawing model's inference layer generates the basic images of each layer sequentially based on the core drawing inference parameter set. This makes the visual hierarchy division of the drawing conform to professional art standards, improving the visual hierarchy expression of AI drawing works and the professionalism and orderliness of layer generation.

[0121] After generating a basic set of layered drawing layers, a feature transfer algorithm is used to perform multi-style fusion processing on each independent layer. This achieves accurate integration and fusion of different art styles on each layer, avoiding style conflicts and abrupt fusion boundaries, and improving the naturalness of style fusion and overall visual unity of multi-style drawing works.

[0122] By extracting brushstroke feature rules from a professional feature rule set, selecting suitable features from a hand-drawn brushstroke feature library, and integrating them into a neural rendering algorithm, the basic drawing image is rendered. This allows AI-generated drawings to combine the accuracy of algorithm generation with the artistic quality of hand-drawn art, improving the artistic expression and visual richness of the drawings and meeting the aesthetic requirements of fine art creation.

[0123] By constructing a five-dimensional aesthetic evaluation system to score the initial version of the artwork, local adjustment requests are generated for dimensions that are below the threshold. An incremental optimization algorithm is used to make precise local adjustments, avoiding the need to redraw the entire image. This improves the artistic professionalism of the final artwork and outputs an editable parameter file, thereby improving the efficiency of subsequent secondary optimization of the artwork.

[0124] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0125] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An AI model-based drawing method, characterized by, Includes the following steps: Step 1: Obtain multimodal drawing requirement data, and use a preset art domain ontology library to convert the multimodal drawing requirement data into structured drawing requirement data; Step 2: Use a feature similarity matching algorithm to match the structured drawing requirement data with various rules in the preset art professional rule base to generate a professional feature rule set; Step 3: Based on the professional feature rule set, select the corresponding parameter adjustment strategy from the preset style-parameter adaptation strategy library to dynamically quantify and assign values ​​to the core inference parameters of the AI ​​drawing model, thereby generating the drawing core inference parameter set of the AI ​​drawing model. Step 4: The AI ​​drawing model divides the model into independent layers according to the professional feature rule set, and generates the base images of each independent layer in sequence according to the drawing core inference parameter set, thus obtaining a layered drawing base layer set. Step 5: Use a seamless layer fusion algorithm to stitch together the layered drawing base layer set, and then perform neural rendering of the stitched drawing base image with hand-drawn brushstrokes to generate a first-version complete drawing work. Step six: Use an incremental optimization algorithm to adjust the local parameters of the initial complete drawing to generate the final drawing and the corresponding editable drawing parameter file.

2. The drawing method based on an AI model according to claim 1, characterized in that, In step one, the multimodal drawing requirement data includes text-based drawing requirement data and visual drawing requirement data; Semantic feature extraction, keyword clustering, and sentiment feature extraction are performed on the text-based drawing requirement data to obtain a text-based requirement feature vector containing a semantic feature vector, a set of core requirement keywords, and sentiment feature values. The visual drawing requirement data is subjected to feature point extraction, contour feature clustering, color histogram extraction, and visual style feature extraction respectively, to obtain a visual requirement feature vector containing visual feature point set, contour feature clustering result, color histogram, and visual style feature; A preset art domain ontology library is loaded. The art domain ontology library contains several core art dimensions. Each core art dimension has a standard feature vector library. The cosine similarity algorithm is used to match the text-type requirement feature vector and the visual-type requirement feature vector with the standard feature vector in the standard feature vector library one by one. The standard feature vector with the highest matching degree is selected for each core art dimension and transformed into the corresponding structured drawing requirement data.

3. The drawing method based on an AI model according to claim 1, characterized in that, In step two, the dimensional rule set of the art professional rule base corresponds to the core art dimensions of the art domain ontology in step one; The feature similarity matching algorithm is used to match the structured drawing requirements data of each core art dimension with the various rules of the corresponding dimension rule set in the preset art professional rule library, and to select a preliminary art rule set; A lightweight embedding algorithm is used to transform the preliminary art rule set into rule parameters that can be recognized by the AI ​​drawing model, and an association index between the rule parameters and the inference layer of the AI ​​drawing model is established to generate a professional feature rule set with a mapping relationship.

4. The drawing method based on an AI model according to claim 1, characterized in that, In step three, core style data is extracted from the professional feature rule set, including art style type, image detail generation requirements, multi-style feature fusion requirements, and light and shadow layer requirements. Based on the art style type, retrieve the corresponding style parameter adjustment strategy from the preset style-parameter adaptation strategy library; Based on the parameter adjustment strategy, and in conjunction with the requirements for generating image details, the requirements for fusion of multiple style features, and the requirements for light and shadow layers, the core inference parameters of the AI ​​drawing model are dynamically quantified and assigned values ​​to generate the drawing core inference parameter set of the AI ​​drawing model. The core inference parameters include generation step size, feature fusion weight, style transfer strength, and detail generation threshold.

5. The AI model-based drawing method of claim 1, wherein, In step four, visual hierarchy rules are extracted from the set of professional feature rules, and the AI ​​drawing model divides the model into several independent layers according to the visual hierarchy rules. The AI ​​drawing model's inference layer generates the base images of each independent layer sequentially based on the core drawing inference parameter set, thus obtaining a layered drawing base layer set.

6. The AI model-based drawing method of claim 1, wherein, In step four, after obtaining the layered drawing base layer set, the feature transfer algorithm is used to perform multi-style fusion processing on the layered drawing base layer set.

7. The AI model-based drawing method of claim 1, wherein, In step five, the neural rendering that integrates hand-drawn brushstrokes includes: Extract brushstroke feature rules from the professional feature rule set, and select suitable hand-drawn brushstroke features from the preset hand-drawn brushstroke feature library according to the brushstroke feature rules; By integrating the characteristics of hand-drawn brushstrokes into the neural rendering algorithm, the basic drawing image obtained from the splicing process is rendered to generate the first version of the complete drawing.

8. The AI model-based drawing method of claim 1, wherein, In step six, a multi-dimensional aesthetic evaluation system is constructed, including the rationality of color matching, the standardization of composition, the accuracy of perspective, the visual balance, and the style fit. The initial complete drawing is scored based on the aforementioned multi-dimensional aesthetic evaluation system. If the score of any dimension in the multi-dimensional aesthetic evaluation system is lower than the preset qualified threshold, a corresponding local parameter adjustment requirement is generated. An incremental optimization algorithm is used to adjust the local parameters of the initial complete drawing based on the local parameter adjustment requirements, generating the final drawing and the corresponding editable drawing parameter file.

9. An AI model-based drawing system employing an AI model-based drawing method according to any one of claims 1 to 8. include: The module includes: drawing requirements analysis module, feature rule mapping module, model parameter generation module, layered layer generation module, drawing screen rendering module, and drawing work optimization module. The drawing requirement parsing module is configured to: acquire multimodal drawing requirement data and convert the multimodal drawing requirement data into structured drawing requirement data using a preset art domain ontology library; The feature rule mapping module is configured to: use a feature similarity matching algorithm to match the structured drawing requirement data with various rules in a preset art professional rule library to generate a professional feature rule set; The model parameter generation module is configured to: based on the professional feature rule set, select the corresponding parameter adjustment strategy from the preset style-parameter adaptation strategy library to dynamically quantify and assign values ​​to the core inference parameters of the AI ​​drawing model, thereby generating the drawing core inference parameter set of the AI ​​drawing model; The layer generation module is configured such that the AI ​​drawing model divides independent layers according to the professional feature rule set, and generates the basic images of each independent layer in sequence according to the drawing core inference parameter set, thereby obtaining a layered drawing basic layer set. The drawing rendering module is configured to: use a seamless layer fusion algorithm to stitch together the layered drawing base layer set, and perform neural rendering that blends hand-drawn brushstrokes on the stitched drawing base image to generate a first-version complete drawing work. The drawing optimization module is configured to use an incremental optimization algorithm to adjust local parameters of the initial complete drawing, generating the final drawing and the corresponding editable drawing parameter file.

Images