A large scene-oriented floor plan and three-dimensional model generation method

Abstract

This invention discloses a method for generating planar layout diagrams and 3D models for large-scale scenes, relating to the fields of 3D vision and layout diagram generation. The invention proposes a hierarchical scene layout generation method that explicitly decouples room topology and geometry to achieve planar layout generation for large scenes. It designs a "2D layout generation + 3D scene transformation" technical route to complete the generation of 3D models for large scenes; it proposes a size-aware semantic centroid heatmap representation, fusing room location and size information predicted by a graph transformer to ensure strong semantic alignment between the generated results and input constraints. It realizes the entire process of planar diagram vectorization and 3D transformation, achieving high-fidelity generation from 2D layout to 3D scene, and uses a large language model to realize the transformation of intelligent furniture layout and natural language description to scene diagram. This method can handle large scenes with more than ten rooms and is applicable to fields such as building information modeling, extended reality, and virtual emergency training.

Description

Technical Field

[0001] This invention belongs to the field of 3D vision and layout generation technology, and relates to a method for generating planar layout diagrams and 3D models for large-scale scenes. Background Technology

[0002] Large-scale indoor 3D scene models are a core technological support for fields such as Building Information Modeling (BIM), Extended Reality (ARP) games, and virtual emergency training. The quality of their generation directly determines the spatial credibility and interactive immersion of the virtual system. Currently, the industry mainly relies on manual production when building large-scale indoor scene models, which is inefficient. While generative artificial intelligence provides a new path for the automated generation of 3D scenes, existing technologies still have significant scale limitations and are difficult to adapt to the generation needs of large-scale scenes.

[0003] Directly generating 3D indoor scene models faces multiple technical challenges: On the one hand, due to the large amount of single sample data, most direct generation methods achieve dimensionality reduction modeling through compressed representation, resulting in the generation results lacking texture details, high inference latency, and poor controllability in large scenes, making it difficult to guarantee semantic alignment; on the other hand, some methods rely on representation techniques such as neural radiation fields, which are prone to causing a lack of geometric consistency from multiple perspectives, producing artifacts and structural distortions, and failing to meet the requirements of large scenes for complex geometric relationships and rich details.

[0004] Given the limitations of direct generation methods, "2D layout generation + 3D scene transformation" has become a more feasible technical approach, often leveraging large language models to generate room layouts and furniture placements. However, existing methods have significant shortcomings in the 2D layout generation stage: layouts directly generated by large language models are too simple, lacking hierarchical processing of structural information in large scenes, making it difficult to represent the complex topological relationships in real large scenes, resulting in insufficient scene realism and becoming a key bottleneck restricting the quality of the final 3D scene.

[0005] Therefore, a method for generating planar layout diagrams and 3D models for large-scale scenes is proposed to solve the above problems. Summary of the Invention

[0006] (a) The technical problem to be solved by the present invention:

[0007] The purpose of this invention is to address the problems of existing methods for directly generating 3D scenes being unable to handle large scene models, and existing 2D planar graph generation methods being unable to generate semantically consistent planar layouts in large scenes. This invention provides a method for generating indoor scene layouts and converting them into 3D scenes for large-scale scenarios. This method explicitly decouples topology and geometry, introduces a size-aware semantic centroid heatmap to ensure strong semantic alignment between input constraints and the generated planar graph; constructs a discrete codebook based on a vector quantization variational autoencoder to represent and generate irregular room shapes; supports concise natural language input to simplify the user interaction process; and designs a complete planar graph vectorization and 3D conversion process to achieve the final 3D scene generation.

[0008] (II) To achieve the above objectives, the present invention adopts the following technical solution: A method for generating planar layout diagrams and 3D models for large-scale scenes includes the following steps: S1. Scene graph construction by fusion of natural language and image: Receives natural language description of a large scene and structural image of the scene from user input, and uses a large language model to convert the natural language description into a scene graph, which includes room nodes and the connection relationships between rooms; S2. Room Attribute Prediction and Heatmap Generation with Topology-Geometric Decoupling: Based on the scene graph and structure image obtained in S1, a topology-geometric explicit decoupling strategy is adopted. A graph transformer is used to predict the centroid position and size of all rooms, and the prediction results are converted into size-aware semantic centroid heatmaps to achieve explicit modeling of the topology. S3. Discrete codebook learning for scene layout based on vector quantization: Construct a discrete codebook representation module and use a vector quantization variational autoencoder to quantize and compress two-dimensional planar layout diagrams containing scene room shapes in a public dataset, and learn their discrete latent representation. S4. Semantic-guided layout generation and door and window layout prediction: Based on the semantic centroid heatmap obtained in S2 and the discrete codebook in S3, a two-dimensional planar layout map is generated using a vector quantization diffusion transformer, and a U-shaped network is used to generate the door and window layout. S5. 3D Scene Reconstruction and Intelligent Furniture Layout Generation: Based on the floor plan generated by S4, vectorization is performed and converted into a 3D scene. A large language model is used to generate the furniture category and placement position for each room. The corresponding furniture model is retrieved from the 3D asset library and placed to obtain a complete 3D interior scene.

[0009] Preferably, the specific implementation process in S1 is as follows: S101. The user inputs a natural language description and a structural image for a large scene. The natural language description specifies the room type, number, and optional adjacency requirements. The structural image represents the spatial outline of the building's load-bearing walls and exterior wall structure, and is used to constrain the boundary conditions of the generated results. S102. The large language model infers the connection relationships between rooms based on statistical constraints and generates a scene graph. ,in Represents a set of semantic room nodes. This represents the set of edges representing the semantic relationships between rooms.

[0010] Preferably, the specific implementation process in S2 is as follows: S201. Use a convolutional neural network as a structural feature extractor to encode structural images into structural features; S202. The scene graph and structural features are input into a multi-layer graph transformer neural network. A multi-head attention mechanism is used to iteratively update node and edge features to capture higher-order spatial relationships and relational dependencies. A cross-attention layer is employed to calculate the dynamic correlation between graph node features and structural features, achieving the fusion of visual and graph features. S203. Decode the node features of the final layer into the room centroid position using the centroid decoder. and room size And based on this, for each room type Generate a single-channel heatmap Each room is represented using a Gaussian kernel centered at the centroid, with a standard deviation proportional to the room size. The specific calculation method is as follows:

[0011] In the formula, The room type is Room collection, With room size Proportional Indicated as the center of mass of the room Axis coordinates; Indicated as the center of mass of the room Axial coordinates; ultimately, all channels are stacked to form a size-aware semantic centroid heatmap.

[0012] Preferably, the iterative update in S202 is... Node features in the layer Sum of edge features The iterative updates of the multi-head attention mechanism are as follows:

[0013]

[0014]

[0015]

[0016] In the formula, , , , , , Represents the learnable projection matrix. This indicates a splicing operation. w Represented as the attention weights between nodes in the scene graph during computation. To focus on the number of heads, for In terms of dimensions, softmax is a normalized exponential function. Finally... and After normalization and feedforward network processing, the following is obtained: and .

[0017] Preferably, the specific implementation process in S3 is as follows: S301. Use a layout encoder to encode the real planar layout graphs in the public dataset into continuous latent tensors, i.e., layout graph features. And maintain a learnable discrete codebook containing 1024 embedding vectors by quantizing each spatial feature vector in the layout graph features into the corresponding entry in the codebook through nearest neighbor search; S302. Collect all quantization vectors to obtain quantization features. The layout is then reconstructed into a planar layout graph using a layout decoder. The layout encoder, layout decoder, and layout graph codebook are jointly optimized using reconstruction loss and quantization loss, enabling them to learn the discrete latent representation of the real planar layout graph.

[0018] Preferably, the corresponding function for the quantization loss in S302 is as follows:

[0019] In the formula, It stops the gradient operation. This is represented as a quantified layout feature.

[0020] Preferably, the specific implementation process in S4 is as follows: S401. Connect the semantic centroid heatmap with the structural features, align it to the latent dimension through convolution operation to form the conditional control tensor of the diffusion process, and add the conditional tensor to the initial noise to form the latent representation of noise. S402. A vector quantization diffusion transformer is used to perform the denoising process. In the denoising iteration, the transformer injects the node features extracted by the graph transformer in S203 through a cross-attention mechanism to better learn the connection and category constraints of the scene graph. S403. After multiple iterative denoising steps, a clean latent representation is obtained. This representation is then used to generate a floor plan layout containing rooms and walls using the layout decoder trained in S302. Finally, a U-shaped network is used to predict the floor plan layout of doors and windows, resulting in a complete 2D scene layout.

[0021] Preferably, the specific implementation process in S5 is as follows: S501. Perform connected component analysis on the generated two-dimensional layout map, extract the outline of each room region, and use the Douglas-Puk algorithm to simplify the outline vertices to obtain the vectorized room outline; and based on the K-nearest neighbor algorithm, use the number of rooms to match similar layout maps in the training dataset, predict the actual scale factor and scale the outline. S503. Extrude each scaled vectorized contour along the vertical axis to form a floor 3D model, extrude the wall contour along the vertical axis to a height H to form a wall 3D model, perform Boolean subtraction operation on it to generate rectangular door and window openings, and accurately place the doors and windows at the preset height to obtain a structural 3D model. S504. For each room, a large language model is used to generate a list of required furniture asset categories, positions, sizes, rotations, and descriptions based on the room type and size. The most matching asset model is then retrieved from the 3D model asset library and placed according to the specified position, size, and rotation. Finally, the furniture is merged with the scene structure 3D model generated in S503 to form a complete 3D interior scene model.

[0022] (III) Compared with the prior art, the layout generation and 3D scene transformation method for large indoor scenes provided by the present invention has the following beneficial effects: (1) This invention proposes a framework for generating planar layout diagrams and three-dimensional models for large-scale scenes. By first generating two-dimensional planar layout diagrams and then converting them into three-dimensional scenes, the technical route avoids the problems of missing details and insufficient controllability in directly generating three-dimensional scene models. It can handle large-scale scene models containing more than ten rooms and realizes semantically consistent and controllable indoor scene generation.

[0023] (2) The present invention designs a hierarchical generation process, which explicitly decouples the topology and geometry of the room in the generation of the planar layout diagram. By predicting the centroid position and size of the room, a size-aware semantic centroid heat map is constructed, which effectively solves the problem of topological complexity in the planar layout diagram of large scenes.

[0024] (3) The present invention uses a vector quantization variational autoencoder to train a discrete codebook to model complex room shapes, and uses a diffusion transformer to generate a high-quality plan view, which effectively solves the problem of generating plan layout diagrams with diverse room shapes.

[0025] (4) This invention provides an automatic conversion method from a two-dimensional layout to a three-dimensional scene, including room vector shape extraction, wall construction, door and window placement and furniture layout, realizing high-fidelity conversion from two-dimensional plan view to three-dimensional scene.

[0026] (5) The present invention achieves optimal performance on large-scale planar layout datasets, while supporting simple natural language input to replace cumbersome graph structure input. The generated three-dimensional scene has high practicality and can be applied to fields such as building information modeling and virtual emergency training. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall process of a method for generating planar layout diagrams and 3D models for large-scale scenes proposed in this invention. Figure 2 This is a schematic diagram of the structure of a dimension-aware semantic centroid heatmap generation module for a method of generating planar layout diagrams and 3D models for large scenes proposed in this invention. Figure 3 This invention presents a method for generating planar layout diagrams and 3D models for large-scale scenes, which generates planar layout diagrams and displays the 3D scene results. Figure 4 This is a comparison chart of the qualitative results of existing mainstream planar layout generation methods with the planar layout diagram and 3D model generation method proposed in this invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1: Please see Figure 1 This invention proposes a method for generating planar layout diagrams and 3D models for large-scale scenes, comprising the following steps: S1. Receive a natural language description and a structural image of a large scene input by the user. Use a large language model to convert the natural language description into a scene graph. The scene graph includes room nodes and the connection relationships between rooms, as follows: S101. The user inputs a natural language description and a structural image for a large scene. The natural language description specifies the room type, number, and optional adjacency requirements. The structural image represents the spatial outline of the building's load-bearing walls and exterior wall structure, and is used to constrain the boundary conditions of the generated results. S102. The large language model infers the connection relationships between rooms based on statistical constraints and generates a scene graph. ,in Represents a set of semantic room nodes. This represents the set of edges representing the semantic relationships between rooms.

[0030] S2. Based on the scene graph and structure image obtained in S1, a topology-geometry explicit decoupling strategy is adopted. A graph transformer is used to predict the centroid location and size of all rooms, and the prediction results are converted into size-aware semantic centroid heatmaps to achieve explicit modeling of the topological structure, as detailed below: S201. Use a convolutional neural network as a structural feature extractor to encode structural images into structural features; S202. The scene graph and structural features are input into a multi-layer graph transformer neural network. A multi-head attention mechanism is used to iteratively update node and edge features to capture higher-order spatial relationships and relational dependencies. A cross-attention layer is employed to calculate the dynamic correlation between graph node features and structural features, achieving the fusion of visual and graph features. The iterative update described in S202 Node features in the layer Sum of edge features The iterative updates of the multi-head attention mechanism are as follows:

[0031]

[0032]

[0033]

[0034] in, , , , , , Represents the learnable projection matrix. This indicates a splicing operation. w Represented as the attention weights between nodes in the scene graph during computation. To focus on the number of heads, for In terms of dimensions, softmax is a normalized exponential function. Finally... and After normalization and feedforward network processing, the following is obtained: and ; S203. Decode the node features of the final layer into the room centroid position using the centroid decoder. and room size And based on this, for each room type Generate a single-channel heatmap Each room is represented using a Gaussian kernel centered at the centroid, with a standard deviation proportional to the room size. The specific calculation method is as follows:

[0035] In the formula, The room type is Room collection, With room size Proportional Indicated as the center of mass of the room Axis coordinates; Indicated as the center of mass of the room Axial coordinates; ultimately, all channels are stacked to form a size-aware semantic centroid heatmap.

[0036] S3. Construct a discrete codebook representation module. Use a vector quantization variational autoencoder to quantize and compress two-dimensional planar layout diagrams containing scene room shapes in a public dataset, and learn their discrete latent representation, as follows: S301. Use a layout encoder to encode the real planar layout graphs in the public dataset into continuous latent tensors, i.e., layout graph features. And maintain a learnable discrete codebook containing 1024 embedding vectors by quantizing each spatial feature vector in the layout graph features into the corresponding entry in the codebook through nearest neighbor search; S302. Collect all quantization vectors to obtain quantization features. The layout is then reconstructed into a planar layout graph using a layout decoder. The layout encoder, layout decoder, and layout graph codebook are jointly optimized using reconstruction loss and quantization loss, enabling them to learn the discrete latent representation of the real planar layout graph.

[0037] The specific function corresponding to the quantization loss described in S302 is as follows:

[0038] in It stops the gradient operation. This is represented as a quantified layout feature.

[0039] S4. Based on the semantic centroid heatmap obtained in S2 and the discrete codebook in S3, a two-dimensional planar layout diagram is generated using a vector quantization diffusion transformer, and a U-shaped network is used to generate the door and window layout, as detailed below: S401. Connect the semantic centroid heatmap with the structural features, align it to the latent dimension through convolution operation to form the conditional control tensor of the diffusion process, and add the conditional tensor to the initial noise to form the latent representation of noise. S402. A vector quantization diffusion transformer is used to perform the denoising process. In the denoising iteration, the transformer injects the node features extracted by the graph transformer in S203 through a cross-attention mechanism to better learn the connection and category constraints of the scene graph. S403. After multiple iterative denoising steps, a clean latent representation is obtained. This representation is then used to generate a floor plan layout containing rooms and walls using the layout decoder trained in S302. Finally, a U-shaped network is used to predict the floor plan layout of doors and windows, resulting in a complete 2D scene layout.

[0040] S5. Based on the floor plan generated in S4, vectorize it and convert it into a 3D scene. Use a large language model to generate the furniture category and placement for each room. Retrieve and place the corresponding furniture models from the 3D asset library to obtain a complete 3D interior scene, as follows: S501. Perform connected component analysis on the generated two-dimensional layout map, extract the outline of each room region, and use the Douglas-Puk algorithm to simplify the outline vertices to obtain the vectorized room outline; and based on the K-nearest neighbor algorithm, use the number of rooms to match similar layout maps in the training dataset, predict the actual scale factor and scale the outline. S503. Extrude each scaled vectorized contour along the vertical axis to form a floor 3D model, extrude the wall contour along the vertical axis to a height H to form a wall 3D model, perform Boolean subtraction operation on it to generate rectangular door and window openings, and accurately place the doors and windows at the preset height to obtain a structural 3D model. S504. For each room, a large language model is used to generate a list of required furniture asset categories, positions, sizes, rotations, and descriptions based on the room type and size. The most matching asset model is then retrieved from the 3D model asset library and placed according to the specified position, size, and rotation. Finally, the furniture is merged with the scene structure 3D model generated in S503 to form a complete 3D interior scene model.

[0041] Example 2: Please see Figure 1-4 Based on Example 1, but with a difference, this invention proposes a method for generating planar layout diagrams and 3D models for large-scale scenes. The specific implementation process is as follows: (a) Data preprocessing: This invention uses the publicly available MSD dataset for training and testing, which contains over 5300 large-scale scene building floor plan layouts. The dataset is divided into training and testing sets according to official standards, and various data augmentation strategies are applied simultaneously to the scene graphs, structure images, and floor plan layouts to enhance the model's generalization ability. These strategies include random rotation, uniform scaling, mirror flipping, and randomly changing certain node types in the scene graphs. (II) Quantitative Compression of Layout Diagrams: Before training the layout graph generation model, a layout graph quantization compression module based on vector quantization variational autoencoder is pre-trained to construct a discrete codebook to represent diverse irregular room shapes. Traditional methods that directly generate complete layouts struggle to effectively model complex non-Manhattan structure rooms. This invention trains this module to compress large-scene planar layouts into a compact latent space, learning a codebook containing 1024 discrete vectors. This codebook captures rich room contour variation patterns, encoding diverse geometric shapes into structured discrete representations, providing high-quality shape priors for subsequent diffusion generation. Training employs a combined optimization strategy of reconstruction loss and quantization loss to ensure the codebook maintains geometric fidelity while possessing good representational compactness. (III) Generation of a hierarchical planar layout diagram guided by the centroid: Building upon the above steps, this invention proposes a centroid-guided hierarchical generation framework to address the dual challenges of generating large-scene planar layout maps by explicitly decoupling topological structure and geometry. In the first stage, a graph transformer fuses scene graph connectivity information and structural image features, simultaneously predicting the centroid positions and sizes of all rooms to construct a size-aware semantic centroid heatmap. This heatmap serves as a topological anchor, ensuring semantic alignment between the generated result and the input constraints. In the second stage, a vector quantization diffusion transformer performs a denoising process in the latent space, guided by the heatmap and discrete codebook, to generate a large-scene planar layout map with complex room shapes. Simultaneously, scene graph information is injected through skip connections to compensate for the limitations of centroid representation. This coarse-to-fine two-stage design effectively reduces generation complexity, enabling the model to handle large scenes containing more than ten rooms. (iv) 3D scene transformation and furniture generation: Based on the generation of a 2D layout, this invention designs a complete process for vectorizing a planar image into a 3D scene, realizing the generation of a 3D scene from a 2D layout. First, a boundary tracking algorithm is used to extract the room outline, and the Douglas-Puk algorithm is used to simplify vertices. The K-nearest neighbor algorithm is used to predict the true scale factor for scale recovery. Then, the vector outline is stretched to generate wall and floor meshes, and Boolean operations are performed to handle door and window openings. Finally, a large language model is used as a virtual interior designer to generate furniture placement based on room semantics, retrieve matching assets from a 3D asset library, and place them in specified poses. This technical approach of generating a 2D layout first and then converting it into a 3D scene effectively avoids the problems of insufficient detail fidelity and poor controllability in methods that directly generate 3D scene models. At the same time, a high-quality 2D layout provides reliable structural constraints for the 3D scene, achieving a balance between generation quality and computational efficiency. like Figure 1 The diagram illustrates the overall flowchart of a method for generating planar layout diagrams and 3D models for large-scale scenes, as proposed in this invention. This method consists of five modules: input and scene graph generation, semantic centroid heatmap generation, layout graph quantization and compression, layout graph and door / window layout generation, and 3D scene generation. In the specific generation process, the user inputs a natural language description and structural image of the large-scale scene. First, a scene graph is generated using a large language model. Then, a graph transformer predicts the room centroids and sizes to construct a semantic centroid heatmap. Next, guided by a discrete codebook, a diffusion transformer generates a planar layout diagram. Finally, after vectorization processing, it is converted into a 3D scene, and furniture is automatically placed. Figure 2 This is a schematic diagram of the structure of the dimension-aware semantic centroid heatmap generation module of the planar layout diagram and 3D model generation method for large-scale scenes proposed in this invention. The graph transformer integrates scene graph connection information and structural image features, and simultaneously predicts the centroid position and size of all rooms. A multi-channel heatmap is constructed through a Gaussian kernel function. This heatmap serves as a topological anchor point to ensure semantic alignment between the generated result and the input constraints, effectively solving the layout chaos problem caused by complex topological relationships in large-scale scenes. Figure 3 This paper presents the results of generating planar layout diagrams and transforming 3D scenes using a method for generating planar layout diagrams and 3D models for large scenes proposed in this invention. It shows four sets of planar layout diagrams generated under different input conditions and their corresponding 3D scene rendering results. The two examples in the first row show pure scene structure 3D models without furniture, containing only building components such as walls, floors, doors, and windows. The two examples in the second row show complete 3D scenes including furniture. It can be seen that this invention can generate large indoor scenes containing multiple room types and complex topological relationships. The 3D transformed scene structure is complete and the furniture layout is reasonable, verifying the effectiveness of this method in practical applications. Figure 4This is a comparison of the qualitative results of the proposed method for generating planar layout diagrams and 3D models for large scenes and the mainstream planar layout diagram generation methods in the prior art. From left to right, the results are the input, the method of the present invention, and the actual value. It can be seen that the existing methods have problems such as room number deviation, connectivity fragmentation, and boundary blurring when extended to large scenes. In contrast, the method of the present invention can more faithfully maintain the structural details and global connectivity at the room level, and the generated planar diagram is highly consistent with the actual value. Table 1 compares the quantitative results of this invention with those of current mainstream planar layout generation methods on the MSD dataset. The quantitative results are evaluated using four metrics: FID, KID, shape similarity, and topological consistency. FID and KID measure the visual quality of the generated image, with lower values ​​being better. Shape similarity measures the similarity of shapes between similar rooms, with higher values ​​being better. Topological consistency measures the degree of connection matching between the generated planar layout and the input scene graph, with higher values ​​being better. Graph2Plan, HouseGAN++, and UN cannot calculate topological consistency because they cannot extract the topological graph of the generated layout. It can be seen that this invention achieves state-of-the-art performance on the MSD dataset.

[0042] Table 1 shows a statistical comparison of the quantitative results of the present invention and existing technologies on the MSD dataset;

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A method for generating planar layout diagrams and 3D models for large-scale scenes, characterized in that, Includes the following steps: S1. Scene graph construction by fusion of natural language and image: Receives natural language description of a large scene and structural image of the scene from user input, and uses a large language model to convert the natural language description into a scene graph, which includes room nodes and the connection relationships between rooms; S2. Room Attribute Prediction and Heatmap Generation with Topology-Geometric Decoupling: Based on the scene graph and structure image obtained in S1, a topology-geometric explicit decoupling strategy is adopted. A graph transformer is used to predict the centroid position and size of all rooms, and the prediction results are converted into size-aware semantic centroid heatmaps to achieve explicit modeling of the topology. S3. Discrete codebook learning for scene layout based on vector quantization: Construct a discrete codebook representation module and use a vector quantization variational autoencoder to quantize and compress two-dimensional planar layout diagrams containing scene room shapes in a public dataset, and learn their discrete latent representation. S4. Semantic-guided layout generation and door and window layout prediction: Based on the semantic centroid heatmap obtained in S2 and the discrete codebook in S3, a two-dimensional planar layout map is generated using a vector quantization diffusion transformer, and a U-shaped network is used to generate the door and window layout. S5. 3D Scene Reconstruction and Intelligent Furniture Layout Generation: Based on the floor plan generated by S4, vectorization is performed and converted into a 3D scene. A large language model is used to generate the furniture category and placement position for each room. The corresponding furniture model is retrieved from the 3D asset library and placed to obtain a complete 3D interior scene.

2. The method for generating planar layout diagrams and 3D models for large-scale scenes according to claim 1, characterized in that, The specific implementation process of scene graph construction in S1, which integrates natural language and images, is as follows: S101. The user inputs a natural language description and a structural image for a large scene. The natural language description specifies the room type, number, and optional adjacency requirements. The structural image represents the spatial outline of the building's load-bearing walls and exterior wall structure, and is used to constrain the boundary conditions of the generated results. S102. The large language model infers the connection relationships between rooms based on statistical constraints and generates a scene graph. ,in Represents a set of semantic room nodes. This represents the set of edges representing the semantic relationships between rooms.

3. The method for generating planar layout diagrams and 3D models for large-scale scenes according to claim 1, characterized in that, The specific implementation process of topological-geometric decoupling room attribute prediction and heatmap generation in S2 is as follows: S201. Use a convolutional neural network as a structural feature extractor to encode structural images into structural features; S202. Input the scene graph and structural features into the multi-layer graph transformer neural network, and iteratively update the node features and edge features through a multi-head attention mechanism to capture higher-order space and relational dependencies; and use a cross-attention layer to calculate the dynamic relationship between graph node features and structural features to achieve the fusion of visual features and graph features. S203. Decode the node features of the final layer into the room centroid position using the centroid decoder. and room size And based on this, for each room type Generate a single-channel heatmap Each room is represented using a Gaussian kernel centered at the centroid, with a standard deviation proportional to the room size. The specific calculation method is as follows: In the formula, The room type is Room collection, With room size Proportional Indicated as the center of mass of the room Axis coordinates; Indicated as the center of mass of the room Axial coordinates; ultimately, all channels are stacked to form a size-aware semantic centroid heatmap.

4. The method for generating planar layout diagrams and 3D models for large-scale scenes according to claim 3, characterized in that, The iterative update described in S202 Node features in the layer Sum of edge features The iterative updates of the multi-head attention mechanism are as follows: In the formula, , , , , , Represents the learnable projection matrix. This indicates a splicing operation. w Represented as the attention weights between nodes in the scene graph; To focus on the number of heads, for In terms of dimensions, softmax is a normalized exponential function; finally... and After normalization and feedforward network processing, the following is obtained: and .

5. The method for generating planar layout diagrams and 3D models for large-scale scenes according to claim 1, characterized in that, The specific implementation process of the scene layout discrete codebook learning based on vector quantization in S3 is as follows: S301. Use a layout encoder to encode the real planar layout graphs in the public dataset into continuous latent tensors, i.e., layout graph features. And maintain a learnable discrete codebook containing 1024 embedding vectors by quantizing each spatial feature vector in the layout graph features into the corresponding entry in the codebook through nearest neighbor search; S302. Collect all quantization vectors to obtain quantization features. The layout is then reconstructed into a planar layout graph using a layout decoder. The layout encoder, layout decoder, and layout graph codebook are jointly optimized using reconstruction loss and quantization loss, enabling them to learn the discrete latent representation of the real planar layout graph.

6. The method for generating planar layout diagrams and 3D models for large-scale scenes according to claim 4, characterized in that, The specific function corresponding to the quantization loss described in S302 is as follows: In the formula, It stops the gradient operation. This is represented as a quantified layout feature.

7. The method for generating planar layout diagrams and 3D models for large-scale scenes according to claim 6, characterized in that, The specific implementation process of semantically guided layout generation and door / window layout prediction in S4 is as follows: S401. Connect the semantic centroid heatmap with the structural features, align it to the latent dimension through convolution operation to form the conditional control tensor of the diffusion process, and add the conditional tensor to the initial noise to form the latent representation of noise. S402. A vector quantization diffusion transformer is used to perform the denoising process. In the denoising iteration, the transformer injects the node features extracted by the graph transformer in S203 through a cross-attention mechanism to better learn the connection and category constraints of the scene graph. S403. After multiple iterations of denoising, a clean latent representation is obtained. The layout decoder trained in S302 generates a planar layout map containing rooms and walls. Then, a U-shaped network is used to predict the planar layout map of doors and windows, and finally, a complete two-dimensional scene layout map is generated.

8. The method for generating planar layout diagrams and 3D models for large-scale scenes according to claim 1, characterized in that, The specific implementation process of 3D scene reconstruction and intelligent furniture layout generation in S5 is as follows: S501. Perform connected component analysis on the generated two-dimensional layout map, extract the outline of each room region, and use the Douglas-Puk algorithm to simplify the outline vertices to obtain the vectorized room outline; and based on the K-nearest neighbor algorithm, use the number of rooms to match similar layout maps in the training dataset, predict the actual scale factor and scale the outline. S503. Extrude each scaled vectorized contour along the vertical axis to form a floor 3D model, extrude the wall contour along the vertical axis to a height H to form a wall 3D model, perform Boolean subtraction operation on it to generate rectangular door and window openings, and accurately place the doors and windows at the preset height to obtain a structural 3D model. S504. For each room, a large language model is used to generate a list of required furniture asset categories, positions, sizes, rotations, and descriptions based on the room type and size. The most matching asset model is then retrieved from the 3D model asset library and placed according to the specified position, size, and rotation. Finally, the furniture is merged with the scene structure 3D model generated in S503 to form a complete 3D interior scene model.

Images