A house type plan generation method based on large model and detection enhancement and thought chain attribute reasoning
Patent Information
- Application Number
- CN202610889989.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-18
- Publication Date
- 2026-10-09
AI Technical Summary
现有方法多采用 “文本→边界框→户型图” 路线，由大模型直接预测房间框体并优化，虽能完成基本生成，但存在明显缺陷：缺乏统一可解释的中间表示，推理过程不透明、难以编辑调试；语义与几何深度耦合，房间数量、邻接、朝向等约束控制能力弱；LLM 输出与扩散模型存在跨模态分布不匹配，易导致结构畸变；未对相对角度、距离等几何属性进行显式推理，稳定性与合规性不足
[0022]与现有技术相比，本发明将大模型的语义理解能力、检索增强与思维链推理机制、气泡图中间表示及扩散模型的几何生成能力有效融合，解决了现有文本驱动户型生成方法语义可控性差、中间表示不可解释、跨模态分布不匹配、几何稳定性不足的问题，具有生成流程可解释、空间结构可控、户型保真度高、适配非专业用户使用等优点，可应用于住宅建筑自动化设计与个性化户型方案快速生成场景。
Smart Images
Figure CN122888066A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent floor plan generation methods, specifically a floor plan generation method based on large model, detection enhancement, and mind chain attribute reasoning. Background Technology
[0002] With the development of building digitization and deep learning technologies, the automated generation of floor plans has become a research hotspot. Traditional floor plan design relies on manual drawing, which suffers from low efficiency, high cost, and difficulty in meeting personalized needs. Furthermore, its limited semantic understanding capabilities make it difficult to interpret abstract design intentions in natural language.
[0003] In recent years, deep learning-based floor plan generation technology has become increasingly mature: early methods based on Generative Adversarial Networks (GANs) could generate basic layouts, but lacked semantic controllability; subsequently, Graph Neural Networks (GNNs) represented floor plans as room relationship diagrams, improving structural rationality, but they were difficult to generate directly from natural language end-to-end; current mainstream diffusion models can generate high-fidelity floor plans, but they struggle to reliably inject text semantic constraints. Bubble charts, as a structured intermediate representation, can clearly express room topology and area relationships, but they are mostly generated manually or according to rules and have not been effectively integrated with large-scale models.
[0004] With the development of large linear modeling (LLM), natural language-driven floor plan generation has become possible. Existing methods mostly adopt a "text → bounding box → floor plan" approach, where the large model directly predicts and optimizes room frames. While this can achieve basic generation, it has significant drawbacks: a lack of unified and interpretable intermediate representations; opaque inference processes that are difficult to edit and debug; deep coupling between semantics and geometry, resulting in weak control over constraints such as room number, adjacency, and orientation; cross-modal distribution mismatch between LLM output and diffusion models, easily leading to structural distortion; and insufficient stability and compliance due to the lack of explicit inference for geometric attributes such as relative angles and distances.
[0005] In summary, existing technologies fail to effectively integrate large models, bubble diagram intermediate representations, retrieval-enhanced thought chains, and diffusion models, making it impossible to achieve semantically accurate, structurally controllable, and geometrically stable text-driven apartment generation, which is insufficient to meet the needs of actual architectural design. Summary of the Invention
[0006] To overcome the shortcomings of existing technologies, this invention provides a method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning. This method achieves a complete process of generating natural language, topological structures, bubble diagrams, and high-precision floor plans, thereby improving semantic controllability, structural rationality, and geometric fidelity.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] A method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning is as follows:
[0009] Step 1: Obtain the user's natural language description text of the apartment type. Use a large model as the topology reasoning module. Semantically analyze the natural language description text of the apartment type through the topology reasoning module to extract room type, number of rooms of each type, room adjacency relationship, and room directional hierarchy relationship. Use each room as a node and the room adjacency relationship as an edge to build a room-level topology graph.
[0010] Step 2: Using the room-level topology map as the query layout, match the query layout with the candidate house type case layouts in the house type case library to obtain the optimal candidate house type case layout.
[0011] Step 3: Use the large model as the thinking chain attribute reasoning module. Use the room-level topology map obtained in Step 1 and the optimal candidate apartment layout obtained in Step 2 as prompts for the large model. Use the large model to perform thinking chain reasoning to predict the center coordinates and area of each room in the room-level topology map. Then, based on the center coordinates and area of each room, obtain the bubble chart corresponding to the room-level topology map.
[0012] Step 4: Using the Diffusion Transformer model, with the bubble chart obtained in Step 3 as a condition and the real floor plan in the training dataset as the target floor plan, the final floor plan corresponding to the room-level topology map is generated through training.
[0013] Furthermore, in step 1, the topology reasoning module determines the room type, room name, and adjacency relationship of each room under predefined architectural prior constraints through semantic parsing; then the topology reasoning module determines the directional hierarchy relationship of the rooms.
[0014] Furthermore, the aforementioned prior architectural constraints include the completion of core rooms, the requirement for the number of adjacent rooms to the bathroom, the bidirectional nature of the adjacency relationship, and the standardization of room naming.
[0015] Furthermore, the topology reasoning module determines the directional hierarchy of rooms, including the parent node of each room and the offset angle range of the child room relative to the parent node.
[0016] Furthermore, in step 2, the Hungarian algorithm is used to match the query layout in the house type case library to obtain several candidate house type case layouts from the house type case library; then the LTSim topological similarity between the query layout and each candidate house type case layout is calculated, and the candidate house type case with the maximum LTSim topological similarity result is obtained as the optimal candidate house type case layout.
[0017] Furthermore, based on room type consistency, orientation angle consistency, parent-child hierarchy consistency, and adjacency consistency, the LTSim topological similarity between the query layout and the layout of each candidate apartment type case is calculated.
[0018] Furthermore, in step 3, the living room in the room-level topology map is taken as the room with the most adjacency relationships. The large model uses the optimal candidate house type layout as a hint to predict the center coordinates of the living room in the room-level topology map, as well as the offset angle and distance of the rooms in the room-level topology map other than the living room relative to the parent node. Then, based on the predicted center coordinates of the living room in the room-level topology map and the offset angle and distance of each room relative to the parent room, the large model predicts the area of each room in the room-level topology map.
[0019] Next, the large model first determines the center coordinates of each room adjacent to the living room in the room-level topology graph using a breadth-first traversal method. Then, it determines the center coordinates of the rooms adjacent to the living room whose center coordinates are not yet determined, thus obtaining the center coordinates of all rooms.
[0020] Furthermore, in step 3, based on the area of each room in the room-level topology map, the area-weighted centroid of all rooms in the room-level topology map is calculated, and combined with the center coordinates of each room, the geometric centroid of the entire apartment layout in the room-level topology map is obtained; based on this geometric centroid, a global translation transformation is performed on the room-level topology map to center-align the apartment layout in the room-level topology map to the center area of the canvas; finally, the large model represents each room in the room-level topology map after the global translation transformation as a circular bubble, with the center of the circle being the center coordinates of the room and the size of the circle being the area of the room, thus obtaining the machine-executable bubble map corresponding to the room-level topology map.
[0021] Furthermore, in step 4, the real floor plans in the training set are used as the target floor plan. The pre-trained VAE encoder of the Diffusion Transformer model is used to encode the target floor plan into a latent space. The bubble diagram embedding is obtained based on the bubble diagram as a condition. The latent space is denoised by the Diffusion Transformer model and then decoded by the VAE to obtain the final floor plan corresponding to the room-level topology map.
[0022] Compared with existing technologies, this invention effectively integrates the semantic understanding capabilities of large models, retrieval enhancement and thought chain reasoning mechanisms, bubble diagram intermediate representation, and the geometric generation capabilities of diffusion models. It solves the problems of poor semantic controllability, uninterpretable intermediate representation, mismatched cross-modal distribution, and insufficient geometric stability in existing text-driven apartment type generation methods. It has the advantages of interpretable generation process, controllable spatial structure, high apartment type fidelity, and suitability for non-professional users. It can be applied to the scenarios of automated design of residential buildings and rapid generation of personalized apartment type schemes. Attached Figure Description
[0023] Figure 1 This is a flowchart of the method according to an embodiment of the present invention. Detailed Implementation
[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0025] like Figure 1 As shown in the figure, this embodiment discloses a method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning. The process is as follows:
[0026] Step 1: Natural Language Input and Room Topology Reasoning
[0027] In this embodiment, the user's natural language description of the apartment layout is obtained. The content of the natural language description text includes the room types in the apartment layout, the number of rooms of each type, and the adjacency relationship of each room. For example, if the user inputs the natural language description text as: "One master bedroom, one secondary bedroom, one living room, one bathroom, one kitchen, one balcony, with the master bedroom located to the upper left of the living room and the balcony adjacent to the master bedroom", then the room types are master bedroom, secondary bedroom, living room, bathroom, kitchen, and balcony, with one room of each type, and the balcony and master bedroom are adjacent to each other.
[0028] In this embodiment, the large model Gpt-4o is used as the topology inference module. The topology inference module performs semantic parsing on the natural language description text of the apartment layout, extracting room types, the number of rooms of each type, room adjacency relationships, and room directional hierarchy relationships. The process is as follows:
[0029] (1) The topology reasoning module determines the room type, room name, and adjacency relationship of each room through semantic parsing under predefined architectural prior constraints. The process is as follows:
[0030] First, the topology reasoning module determines the room type and room name. Then, it prioritizes ensuring the correctness of the adjacency relationships in the user input. Finally, it determines the adjacency relationships between rooms based on the building's prior constraints.
[0031] The building prior constraints include core room completion (e.g., a living room is a necessary room in every house; if the user input does not mention a living room, the topology reasoning module must add the living room when determining the number and type of rooms), the number of adjacent rooms for bathrooms (e.g., a bathroom is generally adjacent to only one room), bidirectional adjacency relationships, and standardized room naming (e.g., if there are two bathrooms, they are named bathroom1 and bathroom2). Adjacency refers to rooms being adjacent to each other through doors.
[0032] (2) The topology reasoning module determines the directional hierarchy of rooms, including the parent node of each room and the offset angle range of the child room relative to its parent node. The process is as follows:
[0033] The topology reasoning module first presets the priority of the rooms, and then selects the room with the highest priority from the adjacent rooms of the rooms other than the living room as the parent node of that room. For example, the master bedroom is connected to the bathroom and the living room, but the living room has a higher priority than the bathroom, so the parent node of the master bedroom is the living room.
[0034] Then, the topology reasoning module, based on the user's input regarding the orientation relationship between the two rooms, converts the orientation relationship of the child nodes into offset angles, obtaining the offset angle range of the child node relative to the parent node. For example, if the master bedroom is located to the upper left of the living room, first determine that the parent node of the master bedroom is the living room, then convert the upper left orientation relationship of the child nodes into offset angles, obtaining the offset angle range of the master bedroom child node relative to the parent node (living room).
[0035] If the user input does not include any directional relationship between rooms, the offset angles of all rooms except the living room and those with directional relationships will be set to None. The specific offset angle values of the rooms other than the living room relative to the parent node will be determined later through the attribute reasoning module.
[0036] Determining the directional hierarchy of rooms allows the subsequently generated bubble chart to satisfy the directional relationships between rooms as input by the user, and facilitates the subsequent attribute reasoning module in determining the offset angle. After obtaining the offset angle, the center coordinates of other rooms besides the living room are determined.
[0037] In this embodiment, a room-level topology graph is finally established, with each room as a node and the adjacency relationships of the rooms as edges.
[0038] Step 2: Enhance case matching
[0039] Using the room-level topology map obtained in step 1 as the query layout, the Hungarian algorithm is used to match the query layout against the apartment type case database to obtain several candidate apartment type case layouts. Then, the LTSim topological similarity between the query layout and each candidate apartment type case layout is calculated, and the candidate apartment type case with the highest LTSim topological similarity is taken as the optimal candidate apartment type case layout.
[0040] In this embodiment, LTSim topological similarity is calculated based on multiple matching dimensions, namely room type consistency, orientation angle consistency, parent-child hierarchy consistency, and adjacency consistency. LTSim Topological Similarity The calculation formula is as follows:
[0041]
[0042] In the formula: Q represents the query layout; C represents the layout for each candidate apartment type case. To query the number of rooms in the layout; In this embodiment, the scaling factor is used. =1; To query the i-th room in layout Q; For each candidate apartment layout matched using the Hungarian algorithm, the rooms are shown in the layout. For single-room matching cost, As shown in the following formula:
[0043]
[0044] In the formula: t represents room type consistency; a represents direction and angle consistency; p represents parent-child hierarchy consistency; and c represents adjacency consistency.
[0045] Room type consistency is calculated using predefined rules: room type consistency for the same room type is 1.0; room type consistency for bedrooms with similar functions is 0.7 or 0.8; room type consistency for kitchens and dining rooms is 0.8; and room type consistency for other unrelated types is 0.1.
[0046] The direction angle consistency is determined by comparing the actual offset angle range of the candidate apartment case room with the offset angle range of the query layout room: if the actual offset angle range of the candidate apartment case room falls within the offset angle range of the query layout room, the direction angle consistency is 1.0; if the actual offset angle range of the candidate apartment case room does not fall within the offset angle range of the query layout room, the direction angle consistency is reduced proportionally according to the ring distance, with the direction angle consistency decreasing as the distance increases, and the minimum direction angle consistency being 0; if the actual offset angle range of the candidate apartment case room is None, None is converted to [0, 360].
[0047] Parent-child hierarchy consistency judgment: If both rooms are root nodes, the parent-child hierarchy consistency is 1.0; if one room is a root node and the other is not, the parent-child hierarchy consistency is 0.3; if neither room is a root node, the similarity is calculated based on the room type of the parent node.
[0048] The adjacency consistency is calculated using Jaccard similarity: the adjacency consistency of two rooms is equal to the size of the intersection of the two sets divided by the size of the union of the two sets, when the two rooms have no adjacency relationship. If the two rooms have no adjacency relationship, the adjacency consistency is recorded as 1.0.
[0049] Therefore, in this embodiment, the optimal candidate unit layout is obtained that is most similar to the topological structure of the room-level topology obtained in step 1. The optimal candidate unit layout provides a true geometric prior for subsequent area allocation and location arrangement, reducing the distribution difference between the large model and the diffusion model.
[0050] Step 3: Reasoning based on the attributes of the thought process
[0051] The large-scale model Gpt-4o is used as the attribute reasoning module for the thought chain. The room-level topology map obtained in step 1 and the optimal candidate apartment layout obtained in step 2 are used as hints for the large-scale model Gpt-4o. The thought chain reasoning is executed step by step through the large-scale model Gpt-4o, as follows:
[0052] First, taking the living room as the room with the most adjacencies in the room-level topology graph, the large model Gpt-4o, using the room-level topology graph and the layout of the optimal candidate apartment type as prompts, predicts the center coordinates of the living room in the room-level topology graph, as well as the offset angles and distances of all rooms in the room-level topology graph relative to their parent nodes. Then, based on the predicted center coordinates of the living room in the room-level topology graph and the offset angles and distances of each room relative to its parent room, the large model Gpt-4o predicts the area of each room in the room-level topology graph.
[0053] Next, the large model Gpt-4o uses a breadth-first search method to first determine the center coordinates of each room adjacent to the living room in the room-level topology graph. Then, it determines the center coordinates of the rooms adjacent to the living room whose center coordinates are not yet determined, thus obtaining the center coordinates of all rooms. The breadth-first search method calculates the room center coordinates using polar coordinate transformation formulas. ),in , , These are the center coordinates of the room's parent node. This represents the offset distance of the current room relative to its parent node. This represents the offset angle of the current room relative to its parent node.
[0054] During the bubble chart generation stage, based on the area of each room in the room-level topology map, the area-weighted centroid of all rooms in the room-level topology map is calculated, and combined with the center coordinates of each room, the geometric centroid of the entire apartment layout in the room-level topology map is obtained. Based on this geometric centroid, a global translation transformation is performed on the room-level topology map to center and align the apartment layout in the room-level topology map to the center area of the canvas, avoiding room coordinates from exceeding the drawing range, ensuring that the bubble chart structure is regular and the position is stable, and providing standardized spatial conditions for the subsequent diffusion model generation.
[0055] Finally, the large model Gpt-4o represents each room in the room-level topology map after global translation transformation as a circular bubble, with the center of the circle being the center coordinate of the room and the size of the circle being the area of the room. This yields a machine-executable bubble map corresponding to the room-level topology map.
[0056] Step 4: Geometric Implementation of the Diffusion Model
[0057] The Diffusion Transformer (DiT) model is used, with the bubble chart obtained in step 3 as a condition and the real floor plan in the training dataset as the target floor plan. The DiT model is trained to generate the final floor plan corresponding to the room-level topology map. The process is as follows:
[0058] The bubble chart is rasterized into a single-channel spatial map, where the position and size of bubbles in different rooms correspond to the room center position and room area, respectively. Convolutional feature extraction is performed on the single-channel spatial map using the conditional image coding network in the DiT model to obtain the bubble chart conditional embedding. The bubble chart conditional embedding includes structural information such as room center position, relative distance, area scale, and spatial topological relationships, which is used to provide global spatial constraints during the diffusion model denoising process.
[0059] Subsequently, the bubble diagram conditional embedding and the diffusion time step embedding are input into the Diffusion Transformer model, enabling the Diffusion Transformer model to perform geometric generation based on the room layout structure expressed by the bubble diagram when predicting noise residuals.
[0060] During the training phase, real floor plans from the training dataset are used as target floor plans. A pre-trained VAE encoder encodes these floor plans into the latent space to obtain target latent variables. Noise is then added to these target latent variables following a diffusion process, creating noisy latent variables. The DiT model takes these noisy latent variables, diffusion time-step embeddings, and bubble chart conditional embeddings as inputs to predict the noise residuals in the noisy latent variables, thereby learning the ability to recover the floor plan latent variables under bubble chart constraints. During the inference phase, the target floor plan is no longer used. Instead, starting from the random noise latent variables, guided by bubble chart conditional embeddings and time-step embeddings, the DiT model progressively denoises the latent variables to obtain the final floor plan latent variables, which are then used by the VAE decoder to generate the final floor plan.
[0061] After denoising, the latent variables are restored to pixel-level floor plans using the VAE decoder in the DiT model; a final floor plan containing walls, doors and windows, room boundaries, reasonable dimensions, and structural specifications is generated.
[0062] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. These embodiments are merely descriptions of preferred embodiments and are not intended to limit the scope or concept of the invention. The specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. Such combinations, as long as they do not violate the spirit of the present invention, should also be considered as part of this disclosure. To avoid unnecessary repetition, the present invention will not further describe the various possible combinations.
[0063] This invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this invention and without departing from the design idea of this invention, all modifications and improvements made by those skilled in the art to the technical solutions of this invention should fall within the protection scope of this invention. The technical content for which protection is sought in this invention has been fully described in the claims.
Claims
1. A method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning, characterized in that, The process is as follows: Step 1: Obtain the user's natural language description text of the apartment type. Use a large model as the topology reasoning module. Semantically analyze the natural language description text of the apartment type through the topology reasoning module to extract room type, number of rooms of each type, room adjacency relationship, and room directional hierarchy relationship. Use each room as a node and the room adjacency relationship as an edge to build a room-level topology graph. Step 2: Using the room-level topology map as the query layout, match the query layout with the candidate house type case layouts in the house type case library to obtain the optimal candidate house type case layout. Step 3: Use the large model as the thinking chain attribute reasoning module. Use the room-level topology map obtained in Step 1 and the optimal candidate apartment layout obtained in Step 2 as prompts for the large model. Use the large model to perform thinking chain reasoning to predict the center coordinates and area of each room in the room-level topology map. Then, based on the center coordinates and area of each room, obtain the bubble chart corresponding to the room-level topology map. Step 4: Using the Diffusion Transformer model, with the bubble chart obtained in Step 3 as a condition and the real floor plan in the training dataset as the target floor plan, the final floor plan corresponding to the room-level topology map is generated through training.
2. The method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning as described in claim 1, characterized in that, In step 1, the topology reasoning module determines the room type, room name, and adjacency relationship of each room through semantic parsing under predefined architectural prior constraints; then the topology reasoning module determines the directional hierarchy relationship of the rooms.
3. The method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning as described in claim 2, characterized in that, The building prior constraints include the completion of core rooms, the number of adjacent rooms for bathrooms, the bidirectional nature of adjacency relationships, and the standardization of room naming.
4. The method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning as described in claim 2, characterized in that, The topology reasoning module determines the directional hierarchy of rooms, including the parent node of each room and the offset angle range of the child room relative to the parent node.
5. The method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning as described in claim 1, characterized in that, In step 2, the Hungarian algorithm is used to match the query layout in the house type case library to obtain several candidate house type case layouts from the house type case library; then, the LTSim topological similarity between the query layout and each candidate house type case layout is calculated, and the candidate house type case with the maximum LTSim topological similarity result is obtained as the optimal candidate house type case layout.
6. The method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning as described in claim 5, characterized in that, Based on room type consistency, orientation angle consistency, parent-child hierarchy consistency, and adjacency consistency, the LTSim topological similarity between the query layout and the layout of each candidate house type case is calculated.
7. The method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning as described in claim 1, characterized in that, In step 3, the living room in the room-level topology map is taken as the room with the most adjacency relationships. The large model uses the layout of the optimal candidate house type case as a prompt to predict the center coordinates of the living room in the room-level topology map, as well as the offset angle and distance of the rooms in the room-level topology map other than the living room relative to the parent node. Then, based on the predicted center coordinates of the living room in the room-level topology map and the offset angle and distance of each room relative to the parent room, the large model predicts the area of each room in the room-level topology map. Next, the large model first determines the center coordinates of each room adjacent to the living room in the room-level topology graph using a breadth-first traversal method. Then, it determines the center coordinates of the rooms adjacent to the living room whose center coordinates are not yet determined, thus obtaining the center coordinates of all rooms.
8. The method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning as described in claim 1, characterized in that, In step 3, based on the area of each room in the room-level topology map, the area-weighted centroid of all rooms in the room-level topology map is calculated, and combined with the center coordinates of each room, the geometric centroid of the entire apartment layout in the room-level topology map is obtained. Based on this geometric centroid, a global translation transformation is performed on the room-level topology map to center and align the apartment layout in the room-level topology map to the center area of the canvas. Finally, the large model represents each room in the room-level topology map after the global translation transformation as a circular bubble, with the center of the circle being the center coordinates of the room and the size of the circle being the area of the room, thus obtaining the machine-executable bubble map corresponding to the room-level topology map.
9. The method for generating floor plans based on large models, detection enhancement, and mind chain attribute reasoning as described in claim 1, characterized in that, In step 4, the real floor plan of the training set is used as the target floor plan. The pre-trained VAE encoder of the Diffusion Transformer model is used to encode the target floor plan into a latent space. The bubble embedding is obtained based on the bubble diagram as a condition. The latent space is denoised by the Diffusion Transformer model and then decoded by VAE to obtain the final floor plan corresponding to the room-level topology map.