Building decoration AI model text map construction method fusing multi-modal data

Abstract

The invention provides a building decoration AI model text map construction method fusing multi-modal data, and belongs to the technical field of building decoration models. Comprising the following steps: S1, collecting multi-modal data, collecting a decoration effect picture, a material texture picture, a spatial layout picture and the like, covering living rooms, bedrooms and kitchens of different styles, modern styles, Chinese styles, European styles and scenes, and arranging design demand description; s2, carrying out standardization processing on the data; s3, cross-modal feature extraction and alignment are carried out; s4, user interaction; and S5, generating a result. According to the method, rich information such as a decoration effect picture, a material texture picture, a spatial layout picture, design requirement description and three-dimensional space structure data can be rapidly collected and integrated by fusing multi-modal data, the data are processed and analyzed by utilizing an advanced AI model, various decoration schemes can be automatically generated, the design period is shortened, and the design efficiency is improved. And the design efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of architectural decoration modeling technology, and in particular to a method for constructing text-based images of architectural decoration AI models that integrate multimodal data. Background Technology

[0002] In the field of architectural interior design, traditional design methods primarily rely on the professional experience and creativity of designers, presenting design schemes through manual sketching and rendering. This process is not only time-consuming and labor-intensive but also inefficient, making it difficult to quickly respond to diverse client needs. While the development of digital technology has led to the emergence of computer-aided design (CAD) software tools, these tools still have limitations in handling complex design requirements, particularly in integrating multiple design elements and achieving personalized customization. In recent years, multimodal data fusion technology has been widely applied in various fields. Multimodal data encompasses different types of data, including images, text, and 3D models, each containing rich information. For example, image data can intuitively display the appearance features of an object, text data can accurately describe the object's attributes and functional requirements, and 3D data can provide information on the object's spatial structure and dimensions. Fusing data from different modalities can fully leverage the advantages of each modality, providing more comprehensive and accurate information support for solving complex problems. The rapid development of artificial intelligence (AI) technology has brought new opportunities to architectural interior design. In particular, deep learning models such as Convolutional Neural Networks (CNNs) and Transformer architectures have achieved remarkable results in areas such as image recognition and natural language processing. Most existing architectural and interior design methods or systems can only process single-modal data and lack the ability to comprehensively process multimodal data. For example, some design software can only design based on two-dimensional images or text descriptions, failing to fully utilize three-dimensional spatial data, resulting in poor applicability of the design results in real-world spaces.

[0003] Therefore, this application provides a method for constructing text-based images of building decoration AI models that integrate multimodal data to meet the requirements. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method for constructing a text-based image of an AI model of building decoration that integrates multimodal data in order to solve the problems of existing methods.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for constructing text-based images for an AI model of building decoration that integrates multimodal data includes the following steps: S1. Multimodal data acquisition: Collect decoration renderings, material texture maps, spatial layout maps, etc., covering different styles, including modern, Chinese, European and scene living rooms, bedrooms, and kitchens. Organize design requirement descriptions, such as "light luxury style, marble background wall, warm color tone", style keyword library, material description documents, and obtain point cloud or grid data of interior space structure walls, door and window positions, furniture dimensions, etc. through laser scanning or CAD models. S2. Standardize the data, unify the image resolution (pixels) and color space (RGB or Lab), segment the text, remove stop words, convert it into word vectors, and embed it using Word2Vec or BERT. 3D data needs to be denoised, registered, and voxelized, and converted into a model-readable format (.obj or .ply). S3. Cross-modal feature extraction and alignment: Using pre-trained convolutional neural networks ResNet and VGG, deep features of the image are extracted, including texture, color distribution, and spatial layout. Transformer architecture, BERT, and GPT are used to encode text descriptions, capturing semantic information such as style, material, and functional requirements. Point cloud networks, PointNet++, or voxel CNNs are used to extract spatial structure features, such as wall height and door and window positions. Different modal features are mapped to the same semantic space through projection matrices or attention mechanisms to ensure cross-modal correlation. The word "marble" in the text corresponds to the texture features in the image. S4. User interaction provides structured input templates such as "Style: [Modern / Chinese]; Material: [Wooden Floor / Marble]; Color: [Warm / Cool]", or free text descriptions. Users can upload hand-drawn sketches or CAD line drawings, extract outlines through image segmentation models, and convert them into generation conditions. It supports uploading indoor scan data, automatically identifies the positions of walls, doors, and windows, and generates decoration schemes that conform to the spatial structure. It can realize partial redrawing to modify the background wall material, style migration to transform modern style into industrial style, and parametric control to adjust furniture size. S5. Generate results. Use models such as ESRGAN to improve image resolution. Use a rule engine to check whether the generated scheme complies with building codes, door and window sizes, and load-bearing structures. Introduce latent spatial interpolation or classifiers to guide sampling and generate multiple variants for users to choose from.

[0006] Preferably, in S1, a high-precision device, such as a LiDAR scanner, needs to be selected, and scanning specifications, including scanning distance and resolution, need to be established. The geometric integrity, closure, and normal direction of the CAD model need to be verified, ultimately forming a full-modal dataset containing images, text, and 3D data.

[0007] Preferably, in S2, the text is tokenized. For Chinese, jieba or THULAC is used to remove stop words such as "的" and "是", and then converted into word vectors. Word2Vec or BERT embedding is used, and a pre-trained model version needs to be selected, either BERT-base or BERT-large. The embedding dimension is determined to be 768. For three-dimensional data, noise reduction, statistical filtering, radius filtering, registration, and ICP algorithm are used to align multi-view scanning data, and voxelization processing is performed.

[0008] Preferably, in S3, a pre-trained convolutional neural network such as ResNet-50 or VGG-16 is used to extract the deep features of the image, including texture, color distribution, and spatial layout. The underlying weights are frozen to retain general features, and only the high-level network, the fully connected layer, is fine-tuned to adapt to the decoration scenario; the Transformer architecture such as BERT or GPT is used to encode the text description to capture semantic information, including style, material, and functional requirements. For short texts, the design requirements, and for long texts, the detailed descriptions, the attention mechanism needs to be optimized separately; a point cloud network such as PointNet++ or a voxel CNN is used to extract spatial structure features such as wall height and door and window positions. The non-uniform point cloud distribution needs to be processed, and the density is balanced through sampling or interpolation. Through a projection matrix, linear transformation, or attention mechanism, cross-modal attention, different modal features are mapped to the same semantic space to ensure cross-modal correlation. For example, the "marble" in the text corresponds to the texture feature in the image. A loss function such as contrastive loss or triplet loss needs to be designed to enhance the feature alignment effect.

[0009] Preferably, in S4, user interaction supports free text descriptions such as "I want a Nordic-style living room with floor-to-ceiling windows". The key information in the free text needs to be parsed through an NLP model such as BERT-NER; users are allowed to upload hand-drawn sketches or CAD line drawings. Through an image segmentation model such as U-Net or Mask R-CNN, the contours are extracted and converted into generation conditions, such as binary masks or edge maps. The segmentation accuracy needs to be optimized, and the recognition rate of hand-drawn sketches is improved through data augmentation. Uploading of indoor scanning data is supported.

[0010] In S5, for the generated results, models such as ESRGAN are used to improve the image resolution, upsampling from 512×512 to 2048×2048. The super-resolution model needs to be optimized for the decoration scenario, and the loss of retaining texture details is increased. The generated scheme is checked through a rule engine to see if it meets the building codes. The door and window sizes need to meet the minimum width requirements of the "Residential Design Code", and the load-bearing structure cannot be modified. Industry knowledge bases such as structural mechanics rules and fire codes need to be embedded, and latent space interpolation is introduced.

[0011] Preferably, the point cloud is converted into a voxel mesh with a resolution of 0.01m, and finally converted into a format readable by the model, .obj or .ply. It is necessary to verify the file format compatibility and check the integrity of vertex and face data.

[0012] Preferably, point cloud or RGB-D images are used to automatically identify the positions of walls, doors, and windows. The plane is fitted using the RANSAC algorithm to generate a decoration scheme that conforms to the spatial structure. Scanning noise needs to be processed, and local redrawing is achieved through filtering or registration optimization. The background wall material needs to be modified, and a masking mechanism needs to be designed. The user can specify the redrawing area. Style transfer is required to transform the modern style into an industrial style. A style encoder and VGG Gram matrix need to be introduced for parameterized control. Furniture size adjustment requires establishing a mapping relationship between parameters and generated results through latent spatial interpolation of conditional GAN.

[0013] Preferably, intermediate styles are generated through linear interpolation, or sampling is guided by a classifier. The generation direction is constrained by the style classifier, which needs to balance diversity. The randomness and correlation of sampling are controlled by temperature parameters. Low-quality results are filtered by threshold. The final output should include high-definition renderings, material lists, construction dimension drawings, etc., and support export in multiple formats, such as JPG, PDF, and DXF.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects: In the above solution, this method can quickly collect and integrate rich information such as decoration renderings, material texture maps, spatial layout maps, design requirement descriptions, and three-dimensional spatial structure data by integrating multimodal data. By using advanced AI models to process and analyze this data, multiple decoration schemes can be automatically generated, which greatly shortens the design cycle and improves design efficiency. At the same time, due to the comprehensive consideration of multiple factors, the generated design schemes are also significantly improved in quality and can better meet the aesthetic and functional needs of users. Users can also upload various forms of information, such as hand-drawn sketches, CAD line drawings, or interior scan data. The system can intelligently analyze this user input, converting it into generation conditions to generate decoration plans that meet the user's personalized needs. Whether it's style preferences, material selection, or spatial layout adjustments, users can customize according to their own wishes, truly achieving personalized design for each individual. Through cross-modal feature extraction and alignment technology, the features of images, text, and 3D data are mapped to the same semantic space, ensuring the correlation and consistency between different modal data. For example, the description of "marble" in the text can accurately correspond to the texture features in the image, ensuring that the generated design plan remains consistent across modalities. This avoids design inconsistencies caused by modal differences and improves the overall integrity and professionalism of the design plan. This contributes to promoting the innovative development of architectural decoration design. Attached Figure Description

[0015] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate embodiments of the invention and, together with the specification, further serve to explain the principles of the invention and enable those skilled in the art to practice and use the invention.

[0016] Figure 1 A schematic diagram illustrating a method for constructing a text-based AI model for building decoration that integrates multimodal data; Figure 2 This is a schematic diagram of data acquisition. Figure 3 This is a diagram illustrating data standardization processing. Figure 4 This is a schematic diagram of feature extraction and alignment. Figure 5 This is a diagram illustrating user interaction. Figure 6 This is a schematic diagram of the generated results.

[0017] As shown in the figure, specific structures and devices are marked in the figure to clearly illustrate the structure of the embodiments of the present invention. However, this is only for illustrative purposes and is not intended to limit the present invention to this specific structure, device and environment. Those skilled in the art can adjust or modify these devices and environments according to specific needs. Detailed Implementation

[0018] The following describes in detail the method for constructing text-based images of an AI model for building decoration that integrates multimodal data, provided by the present invention, with reference to the accompanying drawings and specific embodiments. It should be noted that, to make the embodiments more detailed, the following embodiments are the best and preferred embodiments; those skilled in the art can also use other alternative methods to implement some well-known technologies; and the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0019] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0020] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0021] It is understood that the meanings of “on”, “above”, and “above” in this invention should be interpreted in the broadest manner, such that “on” means not only “directly on” something, but also includes the meaning of being “on” something with an intervening feature or layer, and that “above” or “above” means not only “on” something, but also includes the meaning of being “on” something without an intervening feature or layer.

[0022] Furthermore, spatially related terms such as “below,” “under,” “lower,” “above,” and “upper” are used herein for convenience to describe the relationship of one element or feature to one or more other elements or features, as illustrated in the accompanying drawings. Spatially related terms are intended to cover different orientations in the use or operation of the device other than those depicted in the accompanying drawings. The device may be oriented in other ways, and the spatially related descriptive terms used herein can be interpreted similarly.

[0023] like Figures 1-6 As shown, embodiments of the present invention provide a method for constructing a text-based image of an AI model for building decoration that integrates multimodal data, including the following steps: S1. Multimodal data acquisition: Collect decoration renderings, material texture maps, spatial layout maps, etc., covering different styles, including modern, Chinese, European and scene living rooms, bedrooms, and kitchens. Organize design requirement descriptions, such as "light luxury style, marble background wall, warm color tone", style keyword library, material description documents, and obtain point cloud or grid data of interior space structure walls, door and window positions, furniture dimensions, etc. through laser scanning or CAD models. S2. Standardize the data, unify the image resolution (pixels) and color space (RGB or Lab), segment the text, remove stop words, convert it into word vectors, and embed it using Word2Vec or BERT. 3D data needs to be denoised, registered, and voxelized, and converted into a model-readable format (.obj or .ply). S3. Cross-modal feature extraction and alignment: Use pre-trained convolutional neural networks ResNet and VGG to extract deep features of images, such as texture, color distribution, and spatial layout. Use the Transformer architecture, BERT and GPT, to encode text descriptions and capture semantic information such as style, material, and functional requirements. Use PointNet++ or voxel CNN to extract spatial structure features, such as wall height and door / window positions. Map different modal features to the same semantic space through a projection matrix or attention mechanism to ensure cross-modal relevance. For example, the word "marble" in the text corresponds to the texture feature in the image. S4. User interaction: Provide a structured input template, such as "Style: [Modern / Chinese]; Material: [Wood floor / Marble]; Color: [Warm / Cool]", or free text description. Allow users to upload hand-drawn sketches or CAD line drawings. Extract contours through an image segmentation model and convert them into generation conditions. Support uploading indoor scan data, automatically identify the positions of walls and doors / windows, and generate renovation plans that conform to the spatial structure. Achieve local redrawing, modify the background wall material, style transfer (convert modern style to industrial style), and parametric control to adjust furniture dimensions. S5. Generation results: Use models such as ESRGAN to improve the image resolution. Check whether the generated plan meets building codes, such as door / window sizes and load-bearing structures, through a rule engine. Introduce latent space interpolation or classifier-guided sampling to generate multiple variants for users to choose from.

[0024] In S1, it is necessary to select high-precision equipment, a LiDAR scanner, and formulate scanning specifications, such as scanning distance and resolution. For the CAD model, it is necessary to verify its geometric integrity, closure, and normal direction. Finally, form a full-modal dataset containing images, text, and 3D data.

[0025] In S2, tokenize the text. For Chinese, use jieba or THULAC to remove stop words, such as "de" and "shi", and convert it into word vectors, using Word2Vec or BERT embedding. It is necessary to select a pre-trained model version, BERT-base or BERT-large, and determine the embedding dimension, 768 dimensions. For 3D data, it is necessary to perform noise reduction, statistical filtering, radius filtering, registration, and use the ICP algorithm to align multi-view scan data and voxelization. Convert the point cloud into a voxel grid with a resolution of 0.01m, and finally convert it into a format readable by the model,.obj or.ply. It is necessary to verify the file format compatibility and check the integrity of vertex and patch data.

[0026] In S3, pre-trained convolutional neural networks, such as ResNet-50 and VGG-16, are used to extract deep features of the image, including texture, color distribution, and spatial layout. The weights of the lower layers need to be frozen to retain general features, with only the higher-level networks and fully connected layers fine-tuned to adapt to the decoration scenario. A Transformer architecture, such as BERT and GPT, is used to encode text descriptions, capturing semantic information, style, material, and functional requirements. Design requirements need to be specified for short texts, and detailed for long texts, with attention mechanisms optimized accordingly. Point cloud networks, such as PointNet++ or voxel CNNs, are used to extract spatial structure features, such as wall height and door / window positions. Non-uniform point cloud distribution needs to be addressed by balancing density through sampling or interpolation. Cross-modal attention mechanisms, such as projection matrices, linear transformations, or attention mechanisms, are used to map features from different modalities to the same semantic space, ensuring cross-modal relevance. The term "marble" in the text corresponds to texture features in the image; therefore, loss functions, such as contrastive loss and triplet loss, need to be designed to enhance feature alignment.

[0027] In S4, user interaction supports free text descriptions, such as "I want a Nordic-style living room with floor-to-ceiling windows." This requires using an NLP model, BERT-NER, to parse key information from the free text. Users can upload hand-drawn sketches or CAD line drawings. Image segmentation models, such as U-Net and Mask R-CNN, are used to extract contours and convert them into generating conditions, binary masks, or edge maps. Segmentation accuracy needs optimization, and data augmentation is needed to improve the recognition rate of hand-drawn sketches. Uploading indoor scan data is also supported. Point cloud or RGB-D images are used to automatically identify wall and window locations. A RANSAC algorithm is used to fit a plane and generate a decoration scheme that conforms to the spatial structure. Scan noise needs to be processed, and local redrawing is achieved through filtering or registration optimization. Modifying the background wall material requires a masking mechanism, allowing users to specify the redrawing area. Style transfer, transforming a modern style into an industrial style, requires a style encoder, VGG's Gram matrix, and parameterized control. Adjusting furniture sizes requires establishing a mapping relationship between parameters and generated results, using latent spatial interpolation from conditional GANs.

[0028] In S5, the generated results utilize models such as ESRGAN to upsample the image resolution from 512×512 to 2048×2048. The super-resolution model needs optimization for the decoration scene, increasing the loss for texture detail preservation. A rule engine checks whether the generated scheme complies with building codes; door and window dimensions must meet the minimum width requirements of the "Residential Design Code," and the load-bearing structure cannot be modified. An industry knowledge base, including structural mechanics rules and fire safety codes, must be embedded, and potential spatial interpolation is introduced. Intermediate styles are generated through linear interpolation, or sampling is guided by a classifier. A style classifier constrains the generation direction, balancing diversity. Temperature parameters control sampling randomness and correlation, and thresholds filter low-quality results. The final output must include high-resolution renderings, a material list, and construction dimension drawings, supporting export in multiple formats: JPG, PDF, and DXF.

[0029] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0030] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing text-based images of an AI model for architectural decoration that integrates multimodal data, characterized in that: It includes the following steps: S1. Multimodal data collection: Collect renovation renderings, material texture maps, space layout maps, etc., covering different styles, such as modern, Chinese, European, and scenes like living rooms, bedrooms, and kitchens. Organize design requirement descriptions, such as "luxury style, marble background wall, warm color tone", style keyword libraries, and material description documents. Obtain point cloud or mesh data of the indoor space structure walls, door and window positions, and furniture dimensions through laser scanning or CAD models; S2. Standardize the data: Unify the image resolution pixels and color spaces, RGB or Lab. Tokenize the text, remove stop words, and convert it into word vectors, Word2Vec or BERT embeddings. Three-dimensional data needs to be denoised, registered, and voxelized, and converted into a format readable by the model,.obj or.ply; S3. Cross-modal feature extraction and alignment: Use pre-trained convolutional neural networks ResNet, VGG to extract deep features of images, such as texture, color distribution, and space layout. Use the Transformer architecture, BERT, GPT, to encode text descriptions and capture semantic information such as style, material, and functional requirements. Use point cloud networks, PointNet++, or voxel CNN to extract spatial structure features, such as wall height and door and window positions. Map different modal features to the same semantic space through a projection matrix or attention mechanism to ensure cross-modal relevance. For example, "marble" in the text corresponds to the texture feature in the image; S4. User interaction: Provide a structured input template, "Style: [Modern / Chinese]; Material: [Wood floor / Marble]; Color: [Warm color / Cool color]", or free text description. Allow users to upload hand-drawn sketches or CAD line drawings. Extract the contours through an image segmentation model and convert them into generation conditions, support uploading indoor scan data, automatically identify the positions of walls and doors and windows, generate a renovation plan that conforms to the spatial structure, and implement local redrawing to modify the background wall material, style transfer from modern style to industrial style, and parametric control to adjust furniture dimensions; S5. Generate results: Use models such as ESRGAN to improve the image resolution. Check whether the generated plan meets building codes, such as door and window sizes and load-bearing structures, through a rule engine. Introduce latent space interpolation or classifier-guided sampling to generate multiple variants for users to choose from.

2. The method for constructing a text-based image of an AI model for building decoration that integrates multimodal data according to claim 1, characterized in that, In S1, it is necessary to select high-precision equipment, a LiDAR scanner, and formulate scanning specifications, such as scanning distance and resolution. For CAD models, it is necessary to verify their geometric integrity, closeness, and normal direction. Finally, a full-modal dataset containing images, text, and three-dimensional data is formed.

3. The method for constructing a text-based image of an AI model for building decoration that integrates multimodal data according to claim 1, characterized in that, In S2, tokenize the text. For Chinese, use jieba or THULAC, remove stop words, such as "de" and "shi", and convert it into word vectors, Word2Vec or BERT embeddings. It is necessary to select a pre-trained model version, BERT-base or BERT-large, and determine the embedding dimension, 768 dimensions. Three-dimensional data needs to be denoised, using statistical filtering, radius filtering, registration, the ICP algorithm to align multi-view scan data, and voxelization processing.

4. The method for constructing a text-based image of an AI model for building decoration that integrates multimodal data according to claim 1, characterized in that, In S3, pre-trained convolutional neural networks, ResNet-50 and VGG-16, are used to extract deep features of the image, such as texture, color distribution, and spatial layout. The weights of the bottom layer need to be frozen to retain general features, and only the high-level network and fully connected layers are fine-tuned to adapt to the decoration scenario. Using the Transformer architecture, BERT, and GPT, text descriptions are encoded to capture semantic information, style, material, and functional requirements. Design requirements need to be tailored for short texts and detailed for long texts, with optimized attention mechanisms for each. Point cloud networks, such as PointNet++ or voxel CNNs, are used to extract spatial structural features, such as wall height and door / window positions. Non-uniform point cloud distribution needs to be addressed by balancing density through sampling or interpolation. Cross-modal attention mechanisms, such as projection matrices, linear transformations, or attention mechanisms, are used to map features from different modalities to the same semantic space, ensuring cross-modal relevance. The word "marble" in the text corresponds to texture features in the image; loss functions, such as contrastive loss and triplet loss, need to be designed to enhance feature alignment.

5. The method for constructing a text-based image of an AI model for building decoration that integrates multimodal data according to claim 1, characterized in that, In S4, user interaction supports free text descriptions, such as "I want a Nordic-style living room with floor-to-ceiling windows." This requires using an NLP model, BERT-NER, to parse key information from the free text. Users can also upload hand-drawn sketches or CAD line drawings. Image segmentation models, such as U-Net and Mask R-CNN, are used to extract contours and convert them into generation conditions, binary masks, or edge maps. Segmentation accuracy needs to be optimized, and data augmentation can be used to improve the recognition rate of hand-drawn sketches. Uploading indoor scan data is also supported.

6. The method for constructing a text-based image of an AI model for building decoration that integrates multimodal data according to claim 1, characterized in that, In S5, the generated results are improved by using models such as ESRGAN to upsample the image from 512×512 to 2048×2048. The super-resolution model needs to be optimized for the decoration scene, and the loss of texture detail preservation needs to be increased. The rule engine is used to check whether the generated scheme complies with building codes. The door and window sizes must meet the minimum width requirements of the "Residential Design Code". The load-bearing structure cannot be modified. Industry knowledge bases, structural mechanics rules and fire protection codes need to be embedded, and potential spatial interpolation is introduced.

7. The method for constructing a text-based image of an AI model for building decoration that integrates multimodal data according to claim 3, characterized in that, Convert the point cloud to a voxel mesh with a resolution of 0.01m, and finally convert it to a format readable by the model, such as .obj or .ply. It is necessary to verify the file format compatibility and check the integrity of the vertex and face data.

8. The method for constructing a text-based image of an AI model for building decoration that integrates multimodal data according to claim 5, characterized in that, The system automatically identifies the locations of walls, doors, and windows from point cloud or RGB-D images. It then uses the RANSAC algorithm to fit a plane and generate a decoration scheme that conforms to the spatial structure. This requires processing scanning noise and optimizing local redrawing through filtering or registration. Modifying the background wall material requires designing a masking mechanism. Users can specify the redrawing area. Style transfer is needed to transform a modern style into an industrial style, which requires introducing a style encoder, VGG's Gram matrix, and parametric control. Adjusting furniture dimensions requires establishing a mapping relationship between parameters and generated results, using latent spatial interpolation via conditional GANs.

9. The method for constructing a text-based image of an AI model for building decoration that integrates multimodal data according to claim 6, characterized in that, Intermediate styles are generated through linear interpolation, or sampling is guided by a classifier. The generation direction is constrained by the style classifier, and diversity needs to be balanced. The randomness and correlation of sampling are controlled by temperature parameters. Low-quality results are filtered out by threshold. The final output should include high-definition renderings, material lists, construction dimension drawings, etc., and supports export in multiple formats, such as JPG, PDF, and DXF.

Images