A method for generating Chinese landscape paintings based on frequency domain decoupling enhancement

Abstract

A method for generating Chinese landscape paintings based on frequency domain decoupling enhancement, belonging to the field of image generation, firstly utilizes a large language model to perform semantic understanding and optimization of the natural language prompts input by the user, generating high-quality text guidance signals. Subsequently, combined with a semantic guidance retrieval strategy, two types of key reference images are retrieved from a pre-constructed database: one is authentic Chinese landscape paintings, used to guide the overall artistic style of the generated result; the other is a draft of the brushstroke technique, whose high-frequency components are decoupled through frequency domain analysis, explicitly enhancing the microscopic texture details in the generated image. Through the above collaborative mechanism, this invention achieves high-fidelity, style-controllable, and detail-rich Chinese landscape painting image generation.

Description

Technical Field

[0001] This invention belongs to the field of image generation and relates to a method for generating high-fidelity Chinese landscape paintings based on a diffusion model and frequency domain decoupling. Background Technology

[0002] With the development of artificial intelligence technology, deep learning-based image generation technology has achieved remarkable success in the fields of computer vision and digital art. Especially in the field of natural image generation, existing models are now able to create high-quality works based on text descriptions and have shown great potential in the protection and inheritance of digital cultural heritage.

[0003] Chinese landscape painting is an important part of China's traditional cultural heritage. Its uniqueness lies not only in its macroscopic compositional structure but also in its microscopic brushstroke patterns (i.e., "cunfa"). These cunfa are not merely decorative embellishments but also embody specific landforms and the artist's stylistic intentions. To generate images with a specific artistic style, existing technologies (such as conditional diffusion models like IP-Adapter) typically employ the introduction of reference images as visual constraints. This approach, to some extent, improves the stylistic consistency between the generated image and the reference image.

[0004] However, existing technologies face several insurmountable challenges when applied to the generation of Chinese landscape paintings: First, feature entanglement leads to texture distortion. Existing methods primarily perform feature fusion in the spatial domain. When simultaneously referencing global style and micro-texture images, feature entanglement often occurs between the "micro-texture" and "global composition / color" of different reference images. During generation, this entanglement frequently causes irrelevant backgrounds or color distributions in the reference images to be incorrectly transferred, resulting in overly smoothed textures and disrupting the original macroscopic artistic conception of the landscape painting. Second, user interaction presents a professional barrier. Existing models require users to provide precise reference images or professional prompts to generate high-quality images. For non-professional users, understanding and accurately retrieving specific "texture techniques" is extremely difficult, limiting the widespread adoption of this technology in the field of Chinese landscape painting.

[0005] Therefore, there is an urgent need for an image generation method that can effectively decouple image structure and texture, accurately inject high-quality brushstroke details, and lower the barrier to user interaction. This is particularly important for promoting the digital protection, inheritance, and digital art creation of China's traditional cultural heritage. Summary of the Invention

[0006] To address the problem of entanglement between structural and texture features in existing generative models, leading to the loss of detail in brushstrokes and poor generation quality, this invention proposes a Chinese landscape painting generation method based on frequency domain decoupling enhancement. This method employs a multimodal prompting mechanism. First, it utilizes a large language model to semantically understand and optimize the natural language prompts input by the user, generating high-quality textual guidance signals. Then, combined with a semantic guidance retrieval strategy, it retrieves two types of key reference images from a pre-constructed database: one is authentic Chinese landscape paintings, used to guide the overall artistic style of the generated result; the other is brushstroke sketches, whose high-frequency components are decoupled through frequency domain analysis, explicitly enhancing the microscopic texture details in the generated image. Through this collaborative mechanism, this invention achieves high-fidelity, style-controllable, and detail-rich Chinese landscape painting image generation.

[0007] The technical solution adopted by this invention to solve its technical problem is: A method for generating Chinese landscape paintings based on frequency domain decoupling enhancement includes the following steps: Step S1: Construct a dataset of real Chinese landscape paintings and texture strokes for retrieval. By collecting and labeling high-quality image samples from specific art fields, establish a domain-specific database for the analysis and generation of Chinese landscape painting styles. Step S2: Construct a semantically guided reference information acquisition framework. This framework can receive natural language descriptions input by users, parse the descriptions using a large language model, infer and extract the target "texture texture" category label, and generate optimized prompt words. Based on the optimized prompt words, retrieve matching global style reference images, and retrieve the micro-texture reference images with the highest visual compatibility in the candidate subset according to the "texture texture" category label. Step S3: Construct a texture extraction and enhancement module. Utilize a pre-trained image encoder to extract global style features from the global style reference image and spatial feature maps from the micro-texture reference image. Transform these features to the frequency domain. In the frequency domain, use a high-pass filter to isolate high-frequency texture information and filter out low-frequency structure and color information. Inversely transform the filtered high-frequency components back to the spatial domain to generate pure high-frequency texture features. Resample and compress the high-frequency texture features to obtain frequency-aware texture stroke features. Inject these texture stroke features into the global style features to obtain enhanced style features. Step S4: Establish parallel dual-cross attention paths in the U-Net network of the diffusion model; independently modulate and fuse the spatial features of the U-Net using text features and the enhanced style features respectively, and finally generate a Chinese landscape painting image with high-fidelity brushstrokes.

[0008] Furthermore, in step S1, the process of constructing the dataset of real Chinese landscape paintings and the dataset of brushstrokes for retrieval is as follows: Building a dataset of real landscape paintings Based on existing publicly available digital resources, this study collects images of traditional Chinese landscape paintings. Through manual screening and quality assessment, low-quality samples that are blurry, damaged, or do not conform to specific artistic style standards are removed, ensuring that only high-resolution and high-quality images are retained as the foundational data. Given the lack of detailed textual descriptions in the original images, a multimodal large language model is used to perform visual content analysis on each cleaned image, generating detailed textual annotations covering composition, color scheme, and artistic conception. The resulting dataset contains paired information of "high-quality images - detailed textual descriptions," forming a comprehensive dataset of authentic landscape paintings that reflects global stylistic characteristics. Constructing the texture mapping dataset Under the guidance of experts in the field, we extracted textural samples of brushstrokes with clear technical specifications from Chinese landscape painting teaching materials and student drafts, ensuring that the selected samples were free of background noise and irrelevant brushstroke interference. Based on the traditional landscape painting theory system, we subdivided the textural sample samples into ten representative categories, including hemp fiber brushstrokes, folded ribbon brushstrokes, lotus leaf brushstrokes, small axe-cut brushstrokes, swirling cloud brushstrokes, large axe-cut brushstrokes, chaotic firewood brushstrokes, rice-dot brushstrokes, raindrop brushstrokes, and unraveling rope brushstrokes, to cover the main types of brushstrokes. Thus, we constructed a brushstroke dataset containing a mapping relationship between "brushstroke image-brushstroke category label". This dataset aims to provide accurate brushstroke classification references and lay a solid foundation for subsequent research and applications.

[0009] Furthermore, in step S2, the process of constructing a semantically guided reference information acquisition framework is as follows: Using a large language model to analyze the original prompts from user input Perform semantic analysis to infer the most suitable texturing techniques and extract explicit texture tags. The original prompts were optimized into more precise prompts. ; The CLIP text encoder is used to calculate the cosine similarity between the optimized prompt words and images in the global style library. The image with the highest similarity score is then selected as the global style reference image. The calculation formula is as follows:

[0010] in COS(·) represents the independent variable that maximizes the function, and COS(·) calculates the cosine similarity between two vectors. , These represent the text encoder and the image encoder, respectively. Represents a global style image library; Based on the texture tag In the texture mapping dataset, the search space is narrowed down to a specific candidate subset. Then extract the first 5 images and... The image with the highest visual compatibility is used as the set of micro-texture reference images, and the calculation formula is as follows:

[0011] in, This represents the top five images with the highest cosine similarity scores. This represents a global style image library.

[0012] Furthermore, in step S3, the process of extracting high-frequency texture information is as follows: First, the global reference diagram Input a pre-trained CLIP image encoder to extract global style embeddings. ; Next, for the five texture images selected above, the spatial feature map of the penultimate layer of the CLIP encoder is extracted. Where W, H, and C represent the width, height, and channel dimension of the feature map, respectively. Subsequently, a discrete Fourier transform is applied to the spatial feature map, mapping it to the frequency domain to obtain a complex-valued output. Its expression is as follows:

[0013] in, This represents the frequency components obtained through the Discrete Fourier Transform. To eliminate ringing artifacts caused by ideal filters, a Gaussian high-pass filter is applied. The Gaussian high-pass filter is defined as follows to selectively extract high-frequency components: ; in The scaling parameter of the Gaussian kernel standard deviation is used to control the spatial range of the suppressed low-frequency components, thereby achieving effective decoupling of structure and texture while preserving key high-frequency textures.

[0014] In step S3, the process of obtaining enhanced style features is as follows: The filter and spectrum After element-wise multiplication, the enhanced high-frequency texture feature map is mapped back to the spatial domain using the inverse discrete Fourier transform. : ; in, This represents the inverse discrete Fourier transform. This represents element-wise multiplication; subsequently, the five texture stroke diagrams are compared. Perform average pooling to generate high-frequency prototype vectors. This invention introduces a learnable query mechanism that compresses and refines the high-frequency prototype through a high-frequency resampler: ; in This represents the compressed frequency-domain perceived texture stroke features. This represents a set of learnable query vectors. (·) indicates a high-frequency resampler composed of multiple Transformer modules, supporting continuous interaction between the query vector and the frequency domain prototype. For linear projection layers used for key and value projection; Finally, the compressed high-frequency texture features Injected into the global style features, resulting in frequency-domain enhanced fused features: ; in, This represents the final image features after high-frequency texture enhancement, used for conditional guidance in subsequent generation stages.

[0015] Furthermore, the process of step S4 is as follows: Two parallel cross-attention computation paths are established in the U-Net network, enabling textual and image information to independently address the spatial features. Modulation is performed in this architecture to convert the spatial features The query vector Q is mapped and fed into two attention paths; simultaneously, the features of the input text prompts are utilized. And the enhanced style features output in step S3 Projection generates the corresponding key matrix K and value matrix V, where the key matrix is ​​based on the text prompt word features. Projection yields the text key matrix and text value matrix Based on enhanced style features Projection yields the image key matrix and image value matrix ; Subsequently, the cross-attention of text paths and image paths is calculated independently, and the results of the two calculations are weighted and fused to obtain the updated spatial features. The calculation formula is as follows:

[0016] Where Attention(·) represents the standard cross-attention calculation operation. A scaling factor for adjusting the intensity of the image style; by adjusting It can flexibly control the contribution of the texture stroke style to the final generated result; Finally, utilizing the updated spatial features The diffusion model is then used for denoising. After multiple iterations of denoising, the output is a Chinese landscape painting image that meets the user's semantic expectations and has high-fidelity brushstroke details.

[0017] The beneficial effects of this invention are mainly reflected in the following aspects: Firstly, by constructing a texture extraction enhancement module, frequency domain analysis is creatively introduced into the feature space. Utilizing the prior knowledge that "image structure is distributed in low frequencies and texture details are distributed in high frequencies," the microscopic brushstrokes of Chinese landscape painting are successfully decoupled from the macroscopic composition and color. This mechanism ensures the accurate extraction and injection of pure high-frequency brushstroke features, thus effectively solving the feature entanglement problem commonly found in existing generative models. Secondly, by designing a semantically guided reference information acquisition framework, a multimodal large language model is used to perform semantic parsing and prompt word optimization on the user's natural language, effectively bridging the cognitive gap between colloquial descriptions by ordinary users and professional art terminology. This not only significantly reduces the interaction and usage threshold for non-professional users but also ensures the accuracy of the retrieved and matched reference images and the artistic rigor of the final generated results from the source. Furthermore, a parallel dual cross-attention mechanism is adopted in the generation stage of the diffusion model, independently calculating and dynamically fusing text semantic guidance and frequency-enhanced image style guidance, effectively avoiding the masking of specific artistic texture features by text semantic information under a single fusion path. This application effectively overcomes the shortcomings of existing general diffusion models in generating Chinese landscape paintings, such as excessively smooth textures and lack of artistic authenticity. It provides a high-fidelity, style-controllable and highly automated landscape painting generation solution for the digital protection, intelligent inheritance and digital art creation of traditional Chinese cultural heritage. Attached image description: Figure 1 This is a flowchart of the Chinese landscape painting generation method based on frequency domain decoupling enhancement proposed in this application.

[0018] Figure 2 This is a schematic diagram of the semantically guided reference information acquisition process of this application.

[0019] Figure 3 This is a schematic diagram of the texture extraction and enhancement module of this application.

[0020] Figure 4 This is a schematic diagram of parallel double cross-attention in this application. Detailed Implementation

[0021] The present invention will now be further described with reference to the accompanying drawings.

[0022] Reference Figures 1-4A method for generating Chinese landscape paintings based on frequency domain decoupling enhancement includes: Step S1: Construct a dataset of real Chinese landscape paintings and their texturing techniques for retrieval. By collecting and labeling high-quality image samples from specific art fields, a domain-specific database is established for the analysis and generation of Chinese landscape painting styles.

[0023] In a specific embodiment, the process of constructing the dedicated dataset required for retrieval in step S1 is as follows: Step 1.1: Construct a real landscape painting dataset: Based on existing publicly available digital resources, collect a collection of traditional Chinese landscape painting images. Through manual screening and quality assessment, remove low-quality samples that are blurry, damaged, or do not meet specific artistic style standards. Use a multimodal large language model to perform visual content analysis on each cleaned image to generate detailed text annotations covering information such as composition, color tone, and artistic conception. This will establish a dataset containing paired information of "high-quality image - detailed text description", forming a real landscape painting dataset that comprehensively reflects global style characteristics. Step 1.2: Constructing a texturing texture dataset: Under the guidance of domain experts, extract texturing texture image samples with clear technical specifications from Chinese landscape painting teaching materials and student drafts, ensuring that the selected samples are free from background noise and irrelevant brushstroke interference; based on the traditional landscape painting theory system, subdivide the texturing texture image samples into ten representative categories, including hemp fiber texture, folded ribbon texture, lotus leaf texture, small axe-cut texture, swirling cloud texture, large axe-cut texture, chaotic firewood texture, rice dot texture, raindrop texture, and unraveling rope texture, thereby constructing a texturing texture dataset containing a mapping relationship between "texturing texture image - texturing texture category label".

[0024] Step S2: Construct a semantically guided reference information acquisition framework. This framework can receive natural language descriptions input by users, parse the descriptions using a large language model, infer and extract the target "texture texture" category label, and generate optimized prompt words. Based on the optimized prompt words, retrieve matching global style reference images, and retrieve the micro-texture reference images with the highest visual compatibility in the candidate subset according to the "texture texture" category label.

[0025] In a specific embodiment, such as Figure 2 As shown, in step S2, the process of constructing semantically guided reference information acquisition is as follows: Step 2.1: Utilize a large language model to process the original prompts input by the user. Perform semantic analysis to infer the most suitable texturing techniques and extract explicit texture tags. The original prompts were optimized into more precise prompts. ; Step 2.2: Calculate the cosine similarity between the optimized prompt words and images in the global style library using the CLIP text encoder, and retrieve the image with the highest score as the global style reference image. The calculation formula is as follows:

[0026] in COS(·) represents the independent variable that maximizes the function, and COS(·) calculates the cosine similarity between two vectors. , These represent the text encoder and the image encoder, respectively. Represents a global style image library; Step 2.3: Based on the texture tag In the texture mapping dataset, the search space is narrowed down to a specific candidate subset. Then extract the first 5 images and... The image with the highest visual compatibility is used as the set of micro-texture reference images, and the calculation formula is as follows:

[0027] in, This represents the top five images with the highest cosine similarity scores. This represents a global style image library.

[0028] Step S3: Construct a texture extraction and enhancement module. Use a pre-trained image encoder to extract global style features from the global style reference image and extract spatial feature maps from the micro-texture reference image. Transform the map to the frequency domain and use a high-pass filter to isolate high-frequency texture information while filtering out low-frequency structure and color information. Inversely transform the filtered high-frequency components back to the spatial domain to generate pure high-frequency texture features. Resample and compress the high-frequency texture features to obtain frequency-aware texture stroke features. Inject these texture stroke features into the global style features to obtain enhanced style features.

[0029] In a specific embodiment, such as Figure 3 As shown, the process of constructing the texture extraction and enhancement module in step S3 is as follows: Step 3.1: Input the global reference map into the pre-trained CLIP image encoder to extract the global style embedding. For the five texture images selected above, the spatial feature map of the penultimate layer of the CLIP encoder is extracted. Where W, H, and C represent the width, height, and channel dimension of the feature map, respectively; Step 3.2: Process the spatial feature map. By applying the Discrete Fourier Transform and mapping it to the frequency domain, we obtain a complex numerical output. Its expression is as follows: ; in, This represents the frequency components obtained through the Discrete Fourier Transform. To eliminate ringing artifacts caused by ideal filters, this invention employs a Gaussian high-pass filter. The Gaussian high-pass filter is defined as follows to selectively extract high-frequency components: ; in To adjust the scaling parameter of the Gaussian kernel standard deviation, which controls the spatial range of suppressed low-frequency components, thereby achieving effective decoupling of structure and texture while preserving key high-frequency textures, in this embodiment, it is preferable to set... ; Step 3.3: The filter... and spectrum After element-wise multiplication, the enhanced high-frequency texture feature map is mapped back to the spatial domain using the inverse discrete Fourier transform. : ; in, This represents the inverse discrete Fourier transform. This represents element-wise multiplication. Subsequently, the five texture stroke diagrams are analyzed. Perform average pooling to generate high-frequency prototype vectors. This invention introduces a learnable query mechanism that compresses and refines the high-frequency prototype through a high-frequency resampler: ; in This represents the compressed frequency-domain perceived texture stroke features. This represents a set of learnable query vectors. (·) indicates a high-frequency resampler composed of multiple Transformer modules, supporting continuous interaction between the query vector and the frequency domain prototype. For linear projection layers used for key and value projection; Step 3.4: Compress the high-frequency texture features Injected into the global style features, resulting in frequency-domain enhanced fused features: ; in, This represents the final image features after high-frequency texture enhancement, used for conditional guidance in subsequent generation stages.

[0030] Step S4: Establish parallel dual-cross attention paths in the U-Net network of the diffusion model; independently modulate and fuse the spatial features of the U-Net using text features and the enhanced style features respectively, and finally generate a Chinese landscape painting image with high-fidelity brushstrokes.

[0031] In a specific embodiment, the calculation process of the dual cross-attention path in step S4 is as follows: Step 4.1: Extract spatial features of the current layer of the diffusion model denoising U-Net network. , and map it to the query vector Q through linear projection; Step 4.2 In the text cross-attention branch, utilize text cue word features. Projection yields the text key matrix and text value matrix In the image cross-attention branch, the enhanced style features output in step S3 are utilized. Projection yields the image key matrix and image value matrix ; Step 4.3: Independently calculate the cross-attention of the text path and the image path, and then weight and fuse the results of the two to obtain the updated spatial features. The calculation formula is as follows: ; Where Attention(·) represents the standard cross-attention calculation operation. A scaling factor used to adjust the intensity of image style injection. This is achieved by adjusting... It can flexibly control the contribution of the texture stroke style to the final generated result; Step 4.4: Utilize the updated spatial features The diffusion model is then used for denoising. After multiple iterations of denoising, the output is a Chinese landscape painting image that meets the user's semantic expectations and has high-fidelity brushstroke details.

[0032] Specifically, after the above steps, the CunDiff model proposed in this invention can automatically match the most suitable Chinese landscape painting composition template with professional "cunfa" brushstroke details based on simple natural language prompts from the user. Through frequency domain decoupling technology, this method accurately injects microscopic geological textures, such as the hardness of the axe-cut texture stroke and the softness of the hemp-fiber texture stroke, into the generated image without destroying the macroscopic artistic conception of the landscape painting, such as color and composition. This effectively solves the technical problems of traditional general diffusion models generating Chinese paintings that are "form without spirit," have overly smooth textures, and are prone to feature entanglement.

[0033] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for generating Chinese landscape paintings based on frequency domain decoupling enhancement, characterized in that, The method includes the following steps: Step S1: Construct a dataset of real Chinese landscape paintings and texture strokes for retrieval. By collecting and annotating high-quality image samples from the art field, establish a domain-specific database for the analysis and generation of Chinese landscape painting styles. Step S2: Construct a semantically guided reference information acquisition framework. This framework can receive natural language descriptions input by users, parse the descriptions using a large language model, infer and extract the target "texture texture" category label, and generate optimized prompts. Based on the optimized prompts, search for matching global style reference images, and search for the micro-texture reference images with the highest visual compatibility in the candidate subset according to the "texture texture" category label. Step S3: Construct a texture extraction and enhancement module. Use a pre-trained image encoder to extract global style features from the global style reference image and extract spatial feature maps from the micro-texture reference image. Transform the map to the frequency domain and use a high-pass filter to isolate high-frequency texture information while filtering out low-frequency structure and color information. Inversely transform the filtered high-frequency components back to the spatial domain to generate pure high-frequency texture features. Resample and compress the high-frequency texture features to obtain frequency-aware texture stroke features. Inject these texture stroke features into the global style features to obtain enhanced style features. Step S4: Establish parallel dual-cross attention paths in the U-Net network of the diffusion model; independently modulate and fuse the spatial features of the U-Net using text features and the enhanced style features respectively, and finally generate a Chinese landscape painting image with high-fidelity brushstrokes.

2. The method for generating Chinese landscape paintings based on frequency domain decoupling enhancement as described in claim 1, characterized in that, In step S1, the process of constructing the dataset for retrieval is as follows: Constructing a Real Landscape Painting Dataset: Collecting a collection of traditional Chinese landscape painting images, removing low-quality samples, and using a multimodal large language model to perform visual content analysis on each cleaned image to generate text annotations covering composition, color tone, and artistic conception, thus establishing a real landscape painting dataset containing "high-quality image - detailed text description" pairing information; Constructing a Texturing Technique Dataset: Extracting texturing technique samples with clear technical specifications, and subdividing the texturing technique samples into multiple representative categories according to the traditional landscape painting theory system, thus constructing a texturing technique dataset containing a "texturing technique image - texturing technique category label" mapping relationship.

3. A method for generating Chinese landscape paintings based on frequency domain decoupling enhancement as described in claim 2, characterized in that, In step S2, the process of retrieving the matching reference image is as follows: Using a large language model to analyze the original prompts from user input Perform semantic analysis to infer the most suitable texturing techniques and extract explicit texture tags. The original prompts were optimized into more precise prompts. ; The CLIP text encoder is used to calculate the cosine similarity between the optimized prompt words and images in the global style image library. The image with the highest similarity score is then selected as the global style reference image. ; Based on the texture tag In the texture mapping dataset, the search space is narrowed down to a specific candidate subset. The top 5 texture images with the highest visual compatibility with the global style reference image are extracted as the microtexture reference image set. .

4. The method for generating Chinese landscape paintings based on frequency domain decoupling enhancement as described in claim 3, characterized in that, In step S3, the process of extracting high-frequency texture information is as follows: The global reference map is input into a pre-trained CLIP image encoder to extract global style embeddings. Extract spatial feature maps from the images in the texture image set. ; For the spatial feature map By applying the Discrete Fourier Transform and mapping it to the frequency domain, the frequency domain components are obtained. Applying a Gaussian high-pass filter High-frequency components are selectively extracted, and the spatial range of suppressed low-frequency components is controlled by adjusting the scaling parameter of the Gaussian kernel standard deviation.

5. The method for generating Chinese landscape paintings based on frequency domain decoupling enhancement as described in claim 4, characterized in that, In step S3, the process of obtaining enhanced style features is as follows: The filter and spectrum After element-wise multiplication, the enhanced high-frequency texture feature map is mapped back to the spatial domain using the inverse discrete Fourier transform. ; Perform average pooling on the feature maps corresponding to the texture texture image set to generate high-frequency prototype vectors. ; Introduce a set of learnable query vectors Through high-frequency resampling (·) The high-frequency prototype vector is compressed and refined to obtain the compressed frequency-domain perceptual texture stroke features. ; The compressed frequency-domain perceived texture stroke features Inject into global style embedding In the process, the frequency domain enhanced fusion features are obtained. .

6. A method for generating Chinese landscape paintings based on frequency domain decoupling enhancement as described in any one of claims 1 to 5, characterized in that, In step S4, the process of establishing parallel dual-cross attention paths is as follows: Extracting spatial features of the current layer of the diffusion model-denoising U-Net network , and map it to the query vector Q through linear projection; In the text cross-attention branch, text cue word features are utilized. Projection yields the text key matrix and text value matrix In the image cross-attention branch, the enhanced style features output in step S3 are utilized. Projection yields the image key matrix and image value matrix ; The cross-attention of text paths and image paths is calculated independently, and the results of the two calculations are weighted and fused to obtain the updated spatial features. ; Utilizing updated spatial features The diffusion model is then used for denoising. After multiple iterations of denoising, the output is a Chinese landscape painting image that meets the user's semantic expectations and has high-fidelity brushstroke details.

Images