Existing building facade intelligent generation method with selectable elements

Abstract

The invention relates to an element-selectable existing building facade intelligent generation method, which comprises the following steps of: 1) constructing a facade generation data set and a semantic system, and outputting a classification training data set; 2) performing classification type LoRA training by using the training data set output in the step 1), and finally generating a constraint condition controllable generation model; 3) extracting a to-be-transformed building structure line draft, determining a structure line frame, and drawing a color block definition function element area by a user; outputting a structural line draft graph and a semantic color block graph defined by a user; and 4) performing fusion operation on the model generated in the step 2) and the building structure line draft extracted in the step 3), and restoring the potential feature map into a high-definition image of a pixel space, namely a final high-definition facade scheme. According to the element-selectable existing building facade intelligent generation method, computer vision and the generative AI technology can be combined, structural constraint on the existing building facade is achieved, functional elements are flexible and selectable, and the generation result is accurate and controllable.

Description

Technical Field

[0001] This invention relates to a building facade design method, specifically to an intelligent method for generating existing building facades with selectable elements, belonging to the field of intelligent building design technology. Background Technology

[0002] As urban development shifts from "incremental expansion" to "stock renewal," the revitalization and renovation of existing buildings, such as the renovation of old residential communities, the upgrading of street-front shop facades, and the micro-renewal of historical districts, has become a key area of ​​urban renewal. In such projects, the challenge lies in quickly generating facade renovation plans that both conform to the original structural logic and meet the functional needs of modern commercial formats.

[0003] Existing technologies often have the following main limitations: (1) Traditional design mode is inefficient and has a high threshold: Traditional facade renovation design relies on designers to conduct manual surveying, modeling and rendering, which is time-consuming and costly. For a large number of small and medium-sized businesses or non-professional owners, there is a lack of a low-cost, visual and quick auxiliary design tool.

[0004] (2) Existing AIGC generation technology is “uncontrollable”: Although generative artificial intelligence technology represented by Stable Diffusion performs well in image generation, the generation model tends to prioritize artistic rendering and ignore structural accuracy due to its randomness and creative content. That is, it changes the original load-bearing structure or ignores specific functional openings. At the same time, the model lacks an understanding of functional logic and cannot make accurate layouts according to specific business needs, such as the need for large storefronts in retail. The generated solutions are often good-looking but not practical and cannot be implemented.

[0005] (3) Lack of refined element interaction methods: The existing "image-to-image" technology is mostly for overall style transfer, which makes it difficult to realize the independent control and free combination of local elements of the opposite side, such as only modifying doors and windows while retaining the roof. It cannot meet the actual engineering needs of "preserving the structure and updating the components" in the renovation of existing buildings.

[0006] Therefore, in order to solve the above-mentioned technical problems, it is indeed necessary to provide an innovative method for intelligently generating existing building facades with selectable elements, so as to overcome the defects in the prior art. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent method for generating existing building facades with selectable elements. This method combines computer vision and generative AI technology to achieve flexible selection of structural constraints and functional elements for existing building facades, while ensuring accurate and controllable generation results.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for intelligently generating existing building facades with selectable elements, comprising the following process steps: 1) Construct a facade generation dataset and semantic system, train the U-Net large model to achieve large-scale automated facade semantic segmentation; output a classification training dataset containing "original image + semantic mask + text description"; 2) Training a multi-constraint controllable generative model: Using the training dataset output in step 1), perform classification-based LoRA training, adapt and train the ControlNet-Segmentation network; simultaneously adapt the ControlNet control network; finally, output the business style weight file library and the control network; 3) Extract the structural line drawing of the building to be modified, and perform element-level separation and free combination to determine the structural outline. The user draws color blocks to define functional element areas; output the user-defined structural line drawing and semantic color block diagram. 4) Controllable generation of facade based on structure and function: The model generated in step 2) and the architectural structural line drawing extracted in step 3) are fused together. The latent feature map is restored into a high-definition image in pixel space by the decoder of the variational autoencoder, which is the final high-definition facade scheme.

[0009] The optional existing building facade intelligent generation method of the present invention further comprises: step 1) specifically includes: 1-1) Select existing buildings of the target type and acquire orthophotos of the building facades through high-resolution ground photography; each image synchronously records metadata such as the building's actual coordinates, orientation, number of floors, floor height, camera parameters, and business type; 1-2) Preprocess the images acquired in step 1-1), including distortion correction and perspective transformation, correct the side view to the front elevation view, and uniformly crop them to a standard resolution of 512 x 512 or 1024 x 1024 pixels to construct the original image dataset; (1-3) Based on the needs of building structure and business format, a semantic labeling system is constructed, and a unique RGB color code is assigned to each label to form a semantic color mapping table; according to the business format type recorded in the metadata, the original image dataset is divided into N business format facade subsets for subsequent classification LoRA training; 1-4) Based on the semantic labeling system of steps 1-3), at least 100 representative original images are selected using image annotation tools for pixel-level manual semantic annotation to generate an initial semantic segmentation mask; then, based on the above manually annotated data, a U-Net deep learning model is trained. (1-5) Input the remaining large number of original facade images into the U-Net model trained in steps 1-4) for automated semantic segmentation inference, quickly generate a large-scale semantic segmentation mask map; Simultaneously write a text description for each image, with the description format being "[building type] + [facade material]"; Finally, a four-tuple training dataset consisting of "original image + semantic mask + text description + business type label" is formed.

[0010] The existing building facade intelligent generation method of the present invention can be further described as follows: in step 1-2), invalid samples that are blurry or have more than 30% tree obstruction need to be removed.

[0011] The existing building facade intelligent generation method of the present invention can be further described as follows: In steps 1-3), the semantic tags are divided into: first-level structural tags: including walls, roofs, columns, beams, and bases; second-level functional tags: including doors, windows, balconies, and stairs; and third-level business type tags: including shop signs, air conditioning units, and decorative components.

[0012] The existing building facade intelligent generation method of the present invention, which can be selected as an element, is further described as follows: Step 2) specifically includes: 2-1) The Stable Diffusion V1.5 model was selected as the base for generation; 2-2) For the N business type facade subsets divided in steps 1-3), train N independent LoRA modules respectively. After training, the final output is the business type style weight files LoRA_Retail.safetensors and LoRA_Catering.safetensors; 2-3) Configure the ControlNet control network in Stable Diffusion as a spatial constraint adapter; train and fine-tune two key ControlNet models: ControlNet-Segmentation and ControlNet-Linear Art by inputting the original image and the mask map generated in step 1-5). The trained models are then used in step 4).

[0013] The optional intelligent generation method for existing building facades of the present invention further comprises: in step 2-2), during training, setting the learning rate of the training parameters to 1e. -4 up to 5e -5 The training steps are set to 100 steps per image.

[0014] The optional intelligent generation method for existing building facades of the present invention further comprises: step 3) specifically includes: 3-1) Import the existing photos and mask images of the building to be renovated: Use edge detection algorithms to automatically extract the structural wireframes of the building; then, based on the semantic tagging system of steps 1-3), call the OpenCV library in Python to perform feature-level separation on the extracted wireframes, generating multiple independent line drawing layers; 3-2) Use Boolean logic to freely select and combine the required line art layers to generate a structural line art diagram, thereby defining the final structural constraints; 3-3) Above the wireframe layer determined in step 3-2), create a new "semantic color block map" layer. The user selects the "element" color from the semantic color mapping table in step 1-3 according to the business requirements of the renovation. The user defines the functional attributes of each area on the facade by drawing the position, size and shape of the color blocks. 3-4) Use the structural line drawing after combining steps 3-2) as the first control condition; use the semantic color block drawing drawn by the user in step 3-3) as the second control condition; the user inputs text prompts and specifies the desired style and material.

[0015] The existing building facade intelligent generation method of the optional elements of the present invention is further as follows: In step 3-1), the line drawing layers include layer A: overall building outline line drawing; layer B: roof line drawing; layer C: column structure line drawing; layer D: existing window line drawing.

[0016] The optional existing building facade intelligent generation method of the present invention further includes: step 4) specifically: 4-1), Loading constraints for the model generated in step 2): The "structural line drawing" output in step 3-1) is used as the first control condition and input into the ControlNet-Lineart unit; the "semantic color block" output in step 3-3) is used as the second control condition and input into the ControlNet-Segmentation unit; the user-inputted text prompts are used as the third control condition; 4-2) Style Weight Injection: Based on the user-specified transformation requirements in step 3-3), the model automatically loads the corresponding LoRA weight file from the business style weight file library; 4-3) Initialize the image with Gaussian noise, and then perform 20-50 iterative denoising steps by combining text prompts, structural line drawing features and color block semantic features through inverse diffusion operation; during each denoising step, ControlNet dynamically corrects the feature map; after the iteration is completed, the features of the latent space are decoded into pixel-level high-definition facade images through variational autoencoder to complete the synthesis of a high-definition facade scheme.

[0017] The existing building facade intelligent generation method of the present invention can also be: in step 4-2), the inference weight of LoRA is set to 0.6 ~ 0.8.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. The intelligent generation method for existing building facades, with selectable elements, of this invention achieves precise control over both the structure and function of the generated facade, overcoming the technical defects of existing technologies that often result in structural deformation in generated buildings. This invention extracts the structural line drawings of existing buildings through edge detection algorithms and combines this with ControlNet-Lineart for strong constraints, ensuring that the generated facade design strictly adheres to the perspective relationships and load-bearing structure of the original building, thus achieving structural constraints. Simultaneously, this invention achieves pixel-level precise positioning of functional components such as windows and columns through semantic color block drawing and ControlNet-Segmentation constraints, thus achieving functional constraints. This dual constraint mechanism solves the problem of AIGC-generated buildings being "unworkable."

[0019] 2. The intelligent generation method for existing building facades with selectable elements in this invention pioneers an interactive mode of "free combination of element line drawings," significantly improving the flexibility of renovation. Unlike the traditional "one-click generation" black-box mode, this invention innovatively proposes line drawing layer separation and Boolean combination technology based on semantic tags. Users can freely choose to retain "roof + columns" or only "overall outline," much like building with blocks. This advantage allows the method to perfectly adapt to complex site conditions in existing building renovations (such as certain components must be retained while others must be demolished), truly achieving "selectable elements."

[0020] 3. The intelligent generation method for existing building facades with selectable elements in this invention significantly reduces data processing costs and enables rapid customization of styles for specific business formats. Specifically, this invention utilizes the U-Net model to achieve a leap from "small-sample manual annotation" to "large-scale semi-automatic segmentation," greatly reducing the cost of constructing a facade semantic dataset. Simultaneously, this invention employs a classification-based LoRA training strategy, enabling the model to quickly learn specific materials and decorative styles for different business formats (such as catering and retail). This gives the method strong generalization ability and commercial promotion value, allowing for low-cost and rapid adaptation to renovation projects in different cities and with different styles. Attached Figure Description

[0021] Figure 1 This is a flowchart of an intelligent method for generating existing building facades, which is an optional element of the present invention.

[0022] Figure 2 This is the mask image used for semantic segmentation in steps 1-5 of this invention using the U-Net model.

[0023] Figure 3 It is the line drawing layer formed by freely selecting and combining elements in step 3-2) of this invention.

[0024] Figure 4 This is a schematic diagram of the semantic color blocks drawn by the user according to the business requirements of the transformation in step 3-3) of the present invention. Detailed Implementation

[0025] Please refer to the instruction manual appendix. Figure 1 As shown, this invention is an intelligent method for generating existing building facades with selectable elements, which includes the following process steps: 1) Construct a facade generation dataset and semantic system, and train a large U-Net model to achieve automated large-scale facade semantic segmentation. The output is a classification training dataset containing "original image + semantic mask + text description", which will be directly used as the supervision signal for model training in step 2).

[0026] This step is specifically as follows: 1-1) Select existing buildings of the target type and acquire orthophotos of the building facades through high-resolution terrestrial photography. Each image synchronously records metadata such as the building's actual coordinates, orientation, number of floors, floor height, camera parameters, and business type, which are used for subsequent perspective correction and physical dimension mapping.

[0027] 1-2) Preprocessing is performed on the images acquired in step 1-1), including distortion correction and perspective transformation, correcting the side view to a frontal elevation view, and uniformly cropping them to a standard resolution of 512 x 512 or 1024 x 1024 pixels to ensure the consistency of model input, thereby constructing the original image dataset. In this embodiment, invalid samples with blurriness or tree occlusion exceeding 30% need to be removed to ensure high quality of the original image dataset.

[0028] 1-3) Based on the needs of building structure and business format, a semantic tagging system is constructed. In this embodiment, the semantic tagging system is constructed as shown in Table 1: Table 1 The table divides semantic tags into three levels: Level 1 structural tags, including Wall, Roof, Column, Beam, and Base; Level 2 functional tags, including Door, Window, Balcony, and Stairs; and Level 3 business type tags, including Signage, AC Unit, and Decoration.

[0029] Furthermore, a unique RGB color code is assigned to each label to form a semantic color mapping table. Based on the business type recorded in the metadata, the original image dataset is divided into N business type facade subsets, such as "retail facade subset" and "restaurant facade subset", for subsequent classification-based LoRA training.

[0030] (1-4) Based on the semantic labeling system established in steps 1-3), at least 100 representative original images are selected using the image annotation tool (LabelMe) for pixel-level manual semantic annotation, generating an initial semantic segmentation mask. Then, based on the aforementioned manually annotated data, a U-Net deep learning model is trained. This model is used to learn the mapping relationship from the original image to the semantic segmentation mask, using either Dice Loss or Cross Entropy Loss as the loss function.

[0031] (1-5) Input the remaining large number of original facade images into the U-Net model trained in steps 1-4) for automated semantic segmentation inference, and quickly generate a large-scale semantic segmentation mask map (as shown in the attached image). Figure 2 (As shown). The automatically generated mask images are manually checked and corrected to ensure data quality.

[0032] Furthermore, a text description is simultaneously written for each image, in the format of "[Building Type] + [Facade Material]". This results in a training dataset consisting of a four-tuple of "original image + semantic mask + text description + business type label".

[0033] In this step, the U-Net model was used to achieve a leap from "small sample manual annotation" to "large-scale semi-automatic segmentation", which greatly reduced the construction cost of the facade semantic dataset.

[0034] 2) Training a multi-constraint controllable generative model: Using the training dataset output in step 1), perform classification-based LoRA training, adapt the training to the ControlNet-Segmentation network, and simultaneously adapt the ControlNet control network. The final output is a business style weight file library (LoRA) and a control network (ControlNet) for use in step 4).

[0035] This step is specifically as follows: 2-1) The Latent Diffusion Model (LDM) is selected. In this embodiment, the Stable Diffusion V1.5 model is specifically selected as the base generation platform.

[0036] 2-2) For the N business type facade subsets (such as catering and retail) divided in steps 1-3), train N independent LoRA (low-rank adaptation) modules respectively. By training only the low-rank matrix, each LoRA module focuses on learning the unique material texture and component style of the corresponding business type. For example, use the "retail facade subset" to train the LoRA_Retail module and use the "catering facade subset" to train the LoRA_Catering module.

[0037] Furthermore, during training, the learning rate of the training parameters is set to 1e. -4 up to 5e -5 The training scope is set at 100 steps per image, and the training duration can be balanced based on the amount of data in each subset. The training objective is to enable each LoRA module to focus on learning the subtle differences in materials, decorative details, and lighting of its corresponding business facade, thereby achieving refined style customization.

[0038] After training, the final output is the business style weight files LoRA_Retail.safetensors and LoRA_Catering.safetensors.

[0039] In this step, a classification-based LoRA training strategy is adopted, which enables the model to quickly learn specific materials and decoration styles for different business formats (such as catering and retail). This method has strong generalization ability and commercial promotion value, and can quickly adapt to renovation projects of different cities and styles at low cost.

[0040] 2-3) Configure the ControlNet control network in Stable Diffusion as a spatial constraint adapter for spatial constraints. Train and fine-tune two key ControlNet models, ControlNet-Segmentation and ControlNet-Linear, by inputting the original image and the mask image generated in steps 1-5), establishing a generative model architecture with structural and semantic constraints. The trained model is then used in step 4).

[0041] 3) Extract the structural line drawing of the building to be modified, and perform element-level separation and free combination to determine the structural outline. The user draws color blocks to define the functional element areas. Output the user-defined structural line drawing (Condition A) and semantic color block drawing (Condition B). These two drawings will directly serve as the input signals for step 4).

[0042] This step is specifically as follows: 3-1) Import the existing photos and mask images of the building to be renovated: Use the edge detection algorithm (MLSD algorithm) to automatically extract the structural wireframe of the building. Then, based on the semantic tagging system of steps 1-3), call the OpenCV library in Python to perform feature-level separation on the extracted wireframe, generating multiple independent line drawing layers.

[0043] In this embodiment, the line art layers include layer A: Bounding Box and Overall Outline; layer B: Roofline; layer C: Column / BeamFramework; and layer D: Existing Openings.

[0044] 3-2), Use Boolean logic (AND) to freely select and combine the required line art layers (as shown in the attached image). Figure 3 As shown), generate a structural line drawing to define the final structural constraints. For example: if the user selects (Layer A + Layer B + Layer C), the overall framework, roof, and main structure of the building will be preserved, while allowing windows to change freely. If the user selects (Layer A + Layer D), only the overall outline and existing window positions will be preserved, but roof and column details will be allowed to change.

[0045] In this step, the semantic tag-based line drawing layer separation and Boolean combination technology allows users to freely choose to keep "roof + column" or only keep "overall outline" like building blocks. This advantage makes the method perfectly adaptable to the complex site conditions in the renovation of existing buildings (such as some components must be kept and some must be demolished), truly realizing the "selectable elements" mentioned in the title.

[0046] 3-3), above the wireframe layer determined in step 3-2), create a new "semantic color block map" layer (as shown in the attached image). Figure 4 As shown), users select "element" colors from the semantic color mapping table in steps 1-3) based on the business requirements of the renovation. For example, if the business type is "restaurant", the user selects the color representing "transparent window" and draws a large area of ​​color blocks on the first floor. Users define the functional attributes of each area on the facade by drawing the position, size and shape of the color blocks.

[0047] 3-4) Use the structural line drawing combined in step 3-2) as the first control condition (Condition A); use the semantic color block drawing drawn by the user in step 3-3) as the second control condition (Condition B); the user inputs text prompts (Condition C) to specify the desired style and material.

[0048] 4) Controllable generation of facade based on structure and function: The model generated in step 2) and the architectural structural line drawing extracted in step 3) are fused together. The latent feature map is restored into a high-definition image in pixel space using the decoder of the variational autoencoder (VAE), which is the final high-definition facade scheme.

[0049] Step 4) specifically involves: 4-1), Loading constraints for the model generated in step 2): The "structural line drawing" output in step 3-1) is used as the first control condition (Condition A) and input into the ControlNet-Lineart unit. The "semantic color block diagram" output in step 3-3) is used as the second control condition (Condition B) and input into the ControlNet-Segmentation unit. The user-inputted text prompts are used as the third control condition (Condition C).

[0050] In this step, the MLSD algorithm in step 3-1) is used to extract the structural line drawing of the existing building, and ControlNet-Lineart is used for strong constraints to ensure that the generated facade design strictly follows the perspective relationship and load-bearing structure of the original building, thus achieving structural constraints. Through semantic color block drawing and ControlNet-Segmentation constraints in step 3-3, pixel-level precise positioning of functional components such as windows and columns is achieved, thus achieving functional constraints.

[0051] 4-2) Style Weight Injection: Based on the user-specified transformation requirements (e.g., retail) in step 3-3), the model automatically loads the corresponding LoRA weight file (e.g., LoRA_Retail.safetensors) from the business style weight file library. The LoRA inference strength is set to 0.6 ~ 0.8. This range represents the optimal balance between "basic model image quality" and "specific business style." Too high a weight (>0.8) will lead to excessive image texture stacking and artifacts; too low a weight (<0.6) will fail to effectively reflect the characteristics of a specific business style.

[0052] 4-3) Initialize the image using Gaussian noise, then perform 20-50 steps of iterative denoising using reverse diffusion, combined with text prompts, structural line art features, and color block semantic features. During each denoising step, ControlNet dynamically corrects the feature map to ensure that the generated pixels conform to the boundary constraints of the wireframe and the category constraints of the color blocks. This process forcibly changes the denoising direction, requiring the generated pixels to simultaneously conform to the geometric boundaries of the structural line art and the category distribution of the semantic color blocks.

[0053] After the iteration, the features of the latent space are decoded into pixel-level high-definition facade images by a variational autoencoder (VAE), thus completing the synthesis of a high-definition facade scheme.

[0054] If a user is not satisfied with a certain generated local element (such as the sign style), they can keep the wireframe and other color blocks unchanged, and only modify the color of the color block or the prompt text in that area. They can then use the "inpainting" technique to regenerate that specific area until a satisfactory design is obtained.

[0055] The above-described specific embodiments are merely preferred embodiments of this invention and are not intended to limit this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A method for intelligently generating existing building facades with selectable elements, characterized in that... The process includes the following steps: 1) Construct a facade generation dataset and semantic system, train the U-Net large model to achieve large-scale automated facade semantic segmentation; output a classification training dataset containing "original image + semantic mask + text description"; 2) Training a multi-constraint controllable generative model: Using the training dataset output in step 1), perform classification-based LoRA training, adapt and train the ControlNet-Segmentation network; simultaneously adapt the ControlNet control network; finally, output the business style weight file library and the control network; 3) Extract the structural line drawing of the building to be modified, and perform element-level separation and free combination to determine the structural outline. The user draws color blocks to define functional element areas; output the user-defined structural line drawing and semantic color block diagram. 4) Controllable generation of facade based on structure and function: The model generated in step 2) and the architectural structural line drawing extracted in step 3) are fused together. The latent feature map is restored into a high-definition image in pixel space by the decoder of the variational autoencoder, which is the final high-definition facade scheme.

2. The intelligent generation method for existing building facades with selectable elements as described in claim 1, characterized in that... Step 1) specifically refers to: 1-1) Select existing buildings of the target type and acquire orthophotos of the building facades through high-resolution ground photography; each image synchronously records metadata such as the building's actual coordinates, orientation, number of floors, floor height, camera parameters, and business type; 1-2) Preprocess the images acquired in step 1-1), including distortion correction and perspective transformation, correct the side view to the front elevation view, and uniformly crop them to a standard resolution of 512 x 512 or 1024 x 1024 pixels to construct the original image dataset; (1-3) Based on the needs of building structure and business format, a semantic labeling system is constructed, and a unique RGB color code is assigned to each label to form a semantic color mapping table; according to the business format type recorded in the metadata, the original image dataset is divided into N business format facade subsets for subsequent classification LoRA training; 1-4) Based on the semantic labeling system of steps 1-3), at least 100 representative original images are selected using image annotation tools for pixel-level manual semantic annotation to generate an initial semantic segmentation mask; then, based on the above manually annotated data, a U-Net deep learning model is trained. (1-5) Input the remaining large number of original facade images into the U-Net model trained in steps 1-4) for automated semantic segmentation inference, quickly generate a large-scale semantic segmentation mask map; Simultaneously write a text description for each image, with the description format being "[building type] + [facade material]"; Finally, a four-tuple training dataset consisting of "original image + semantic mask + text description + business type label" is formed.

3. The intelligent generation method for existing building facades with selectable elements as described in claim 2, characterized in that... In steps 1-2), invalid samples that are blurry or obscured by trees by more than 30% need to be removed.

4. The intelligent generation method for existing building facades with selectable elements as described in claim 2, characterized in that... In steps 1-3), the semantic tags are divided into: Level 1 structural tags: including walls, roofs, columns, beams, and bases; Level 2 functional tags: including doors, windows, balconies, and stairs; Level 3 business type tags: including shop signs, air conditioner units, and decorative components.

5. The intelligent generation method for existing building facades with selectable elements as described in claim 2, characterized in that... Step 2) specifically refers to: 2-1) The Stable Diffusion V1.5 model was selected as the base for generation; 2-2) For the N business type facade subsets divided in steps 1-3), train N independent LoRA modules respectively. After training, the final output is the business type style weight files LoRA_Retail.safetensors and LoRA_Catering.safetensors; 2-3) Configure the ControlNet control network in Stable Diffusion as a spatial constraint adapter; train and fine-tune two key ControlNet models: ControlNet-Segmentation and ControlNet-Linear Art by inputting the original image and the mask map generated in step 1-5). The trained models are then used in step 4).

6. The intelligent generation method for existing building facades with selectable elements as described in claim 5, characterized in that... In step 2-2), during training, the learning rate of the training parameters is set to Ie. -4 up to 5e -5 The training steps are set to 100 steps per image.

7. The intelligent generation method for existing building facades with selectable elements as described in claim 2, characterized in that... Step 3) specifically refers to: 3-1) Import the existing photos and mask images of the building to be renovated: Use edge detection algorithms to automatically extract the structural wireframes of the building; then, based on the semantic tagging system of steps 1-3), call the OpenCV library in Python to perform feature-level separation on the extracted wireframes, generating multiple independent line drawing layers; 3-2) Use Boolean logic to freely select and combine the required line art layers to generate a structural line art diagram, thereby defining the final structural constraints; 3-3) Above the wireframe layer determined in step 3-2), create a new "semantic color block map" layer. The user selects the "element" color from the semantic color mapping table in step 1-3 according to the business requirements of the renovation. The user defines the functional attributes of each area on the facade by drawing the position, size and shape of the color blocks. 3-4) Use the structural line drawing after combining steps 3-2) as the first control condition; use the semantic color block drawing drawn by the user in step 3-3) as the second control condition; the user inputs text prompts and specifies the desired style and material.

8. The intelligent generation method for existing building facades with selectable elements as described in claim 7, characterized in that... In step 3-1), the line art layers include layer A: overall building outline line art; layer B: roof line art; layer C: column structure line art; and layer D: existing window line art.

9. The intelligent generation method for existing building facades with selectable elements as described in claim 7, characterized in that... Step 4) specifically refers to: 4-1), Loading constraints for the model generated in step 2): The "structural line drawing" output in step 3-1) is used as the first control condition and input into the ControlNet-Lineart unit; the "semantic color block" output in step 3-3) is used as the second control condition and input into the ControlNet-Segmentation unit; the user-inputted text prompts are used as the third control condition; 4-2) Style Weight Injection: Based on the user-specified transformation requirements in step 3-3), the model automatically loads the corresponding LoRA weight file from the business style weight file library; 4-3) Initialize the image with Gaussian noise, and then perform 20-50 iterative denoising steps by combining text prompts, structural line drawing features and color block semantic features through inverse diffusion operation; during each denoising step, ControlNet dynamically corrects the feature map; after the iteration is completed, the features of the latent space are decoded into pixel-level high-definition facade images through variational autoencoder to complete the synthesis of a high-definition facade scheme.

10. The intelligent generation method for existing building facades with selectable elements as described in claim 9, characterized in that... In step 4-2), the inference weight of LoRA is set to 0.6 ~ 0.8.

Images