A knowledge-enhanced and two / three-dimensional linkage characteristic style full-process design method

Abstract

The application relates to a characteristic style full-process design method combining knowledge enhancement and two-dimensional and three-dimensional linkage, which comprises the following steps: acquiring multi-modal data of a target style type and preprocessing; establishing building physical height and image gray mapping criteria and a ground texture morphological index feature library; constructing an expert and public dual-view multi-agent system, using a thinking chain for multi-agent collaborative analysis, and constructing a knowledge graph with design method co-occurrence quantification and feature weight visualization; training a texture patching LoRA model and a knowledge-enhanced facade design LoRA model respectively, and constructing a knowledge-enhanced prompt word engineering; performing full-process two-dimensional and three-dimensional linkage scheme generation, including divergence generation of plane texture, two-dimensional and three-dimensional linkage vectorization conversion, morphological index accounting and inspection screening, two-dimensional and three-dimensional linkage block stretching deduction, and controllable generation of facade style. Compared with the prior art, the application has the advantages of strong design divergence, high professionalism, high efficiency, and higher scientific nature.

Description

Technical Field

[0001] This invention relates to the field of urban landscape design, and in particular to a method for the whole-process design of distinctive landscapes that combines knowledge enhancement and two-dimensional / three-dimensional linkage. Background Technology

[0002] As my country's urbanization shifts from scale expansion to quality improvement, how to preserve the unique character of cities and avoid homogenization amidst intensive urban renewal has become a core issue in urban design and architectural creation. The unique character of a city encompasses not only the visual expression of building facades but also the organizational logic of its planar texture and the volumetric order of its three-dimensional space.

[0003] Traditional urban design methods rely heavily on the personal experience of designers, and when faced with large-scale textural patching and complex volume deduction, they suffer from limitations such as long design cycles, low iteration efficiency, and narrow solution coverage.

[0004] In recent years, the explosive growth of generative artificial intelligence technology has provided an opportunity for the intelligent transformation of urban design, improving creative efficiency and the diversity of solutions to some extent. However, although existing generative technologies have made some progress in the stylized generation of single images, the industry still faces the following common challenges when applied to the entire urban design process:

[0005] First, there is a break in the two-dimensional and three-dimensional linkage of spatial logic. Existing image generation technology is essentially a probabilistic combination of two-dimensional pixels, lacking a deep encoding mechanism based on physical properties. This makes it difficult to establish a precise geometric mapping relationship between the generated facade renderings and planar textures and three-dimensional blocks, failing to achieve a closed-loop linkage from visual presentation to spatial entities.

[0006] Second, the mechanism for injecting professional knowledge is still immature. The general model lacks a professional understanding of the unique characteristics of the city (such as spatial structure and facade design). Without the enhancement of professional knowledge, the generated solutions, while visually appealing, are significantly distorted in terms of architectural logic and cultural core. At the same time, due to the lack of effective integration with the public's emotional needs, the generated results are still insufficient in terms of social acceptability.

[0007] Third, the design process is fragmented. Existing intelligent methods mostly focus on improving efficiency at individual stages, and have not yet formed a progressive closed loop covering planar texture restoration, three-dimensional block construction, and facade image generation. The lack of unified data interfaces and automatic verification mechanisms based on morphological indicators between different design stages makes it difficult to guarantee the feasibility and professional accuracy of the generated solutions.

[0008] Therefore, there is an urgent need for a generative urban design method that can achieve two-dimensional and three-dimensional linkage, cover the entire design process, and deeply integrate multi-dimensional knowledge, so as to promote the professional application of intelligent technology in shaping the urban landscape.

[0009] Chinese invention patent CN117671220A discloses a method for optimizing urban planning implementation based on virtual reality technology. The method includes: real-time collection of built environment data for a target urban area and obtaining planning implementation information; constructing a digital sand table for urban planning implementation in the ArcGIS Pro platform to generate 3D building models; performing slice extraction and virtual reality walkthrough display to showcase the implementation status of key areas; evaluating the implementation plan using morphological and landscape scoring modules and marking slices requiring adjustment; performing real-time adjustments and optimizations using an Oculus Rift VR headset and Leap Motion handheld controller; and integrating the scoring process and results of each slice to generate an urban planning implementation evaluation and optimization report. This invention can track user actions and gestures in real time, correcting building height and floor area ratio information of the implemented plan, and adjusting material texture and color information, thereby completing the optimization and adjustment of the plan. However, there are still problems such as the lack of professional design knowledge guidance in the design process and excessive manual workload, the fragmentation of cross-dimensional semantics leading to extremely low system integration of the whole process design and difficulty in the transformation of results, the rigid and singular output solutions of existing rule-based or instruction-driven technical solutions, and the lack of effective screening mechanisms that make it difficult to verify style consistency from the perspective of urban morphology.

[0010] In summary, there is currently a lack of a comprehensive design methodology that integrates knowledge enhancement and 2D / 3D integration to address or partially resolve the aforementioned issues. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a knowledge-enhanced and two-dimensional linkage-based full-process design method for distinctive landscapes. This method aims to solve or partially solve the problems of lack of professional design knowledge guidance and excessive manual workload in the design process, the low system integration and difficulty in transforming results due to the break in cross-dimensional semantics, the rigid and singular logical convergence output of existing rule-based or instruction-driven technical solutions, and the lack of an effective screening mechanism that makes it difficult to verify style consistency from the perspective of urban morphology.

[0012] The objective of this invention can be achieved through the following technical solutions:

[0013] This invention provides a method for the full-process design of distinctive landscape features that combines knowledge enhancement and 2D / 3D linkage. The method includes:

[0014] S1. Acquire multimodal data of the target landscape type, establish a mapping criterion between building physical height and image grayscale, and based on the mapping criterion between building physical height and image grayscale, perform grayscale preprocessing of height information and standardized sampling unit spatial cutting on the local texture map in the multimodal data, extract morphological indicators of texture samples in the sampling unit, define the confidence interval of local texture typology indicators, and establish a feature library of local texture morphology indicators;

[0015] S2. Construct a multi-agent system with dual perspectives of experts and the public, and use a thought chain to guide the multi-agent system to perform multi-agent collaborative analysis of the multimodal data. Then, summarize and quantify the analysis results, construct a multimodal knowledge graph with element-method-case-image mapping that has co-occurrence quantification of design techniques and visualization of feature weights, and construct a recommendation of high- and low-frequency design technique combinations based on the knowledge graph.

[0016] S3. Train a texture patching LoRA model based on the standardized two-dimensional local texture map obtained by spatial cutting of the standardized sampling unit; train a knowledge-enhanced facade design LoRA model based on the data in the multimodal knowledge graph; and construct a knowledge-enhanced prompt word project based on the combination of high- and low-frequency design techniques.

[0017] S4. Using the trained texture patching LoRA model, combined with local redrawing divergent generation of planar texture, a two-dimensional planar texture bitmap is obtained; the planar texture bitmap is converted into a vector building outline, and according to the building physical height and image grayscale mapping criterion, the grayscale information of the planar texture bitmap is converted into the building physical height attribute of each closed vector building outline. Based on the vector building outline and building physical height attribute, morphological base data is formed, and full-dimensional morphological indicators are calculated according to the morphological base data; the full-dimensional morphological indicators are compared with the confidence interval of the local texture typology indicators in real time, and the schemes with indicator distribution that meet the constraints are selected. Using the vector building outline and the building physical height and image grayscale mapping criterion, two-dimensional and three-dimensional linked block stretching is performed to obtain a three-dimensional building volume model. Combined with the knowledge-enhanced prompt word engineering, the facade style is generated in collaboration with the depth constraint and the trained knowledge-enhanced facade design LoRA model, and the final distinctive style design scheme is output.

[0018] As a preferred technical solution, the multimodal data includes academic theoretical texts, practical case studies and graphic materials, and local texture maps; the local texture map is a vector diagram of the architectural texture carrying local characteristics within the urban area of ​​the target landscape type; the morphological indicators are a set of geometric and statistical parameters that quantitatively describe the spatial organization characteristics of the city; the indicators in the local texture morphological indicator feature library include test indicators for the number of buildings, average number of floors, base area, total building area, plot ratio, building density, and open space ratio.

[0019] As a preferred technical solution, the expert and public dual-perspective multi-agent system is constructed using a large language model, and the specific steps include:

[0020] Construct an expert-perspective intelligent agent system, which possesses architectural expertise, and systematically analyzes the multimodal data to extract professional design patterns.

[0021] A public-perspective intelligent agent system is constructed. Based on the assigned differentiated identity profile, including age, occupation and aesthetic preferences, the public-perspective intelligent agent system simulates the visual aesthetics and interest tendencies of different social groups and identifies visual features in multimodal data that match the interests of the corresponding groups.

[0022] As a preferred technical solution, the multi-agent collaborative analysis of the multimodal data is guided by the mind chain of the expert and public dual-perspective multi-agent system. The specific steps are as follows:

[0023] S2.1. Establish the scale of the parsing task and the criteria for knowledge saturation. When the number of new knowledge extracted is lower than the preset threshold, it is determined that the knowledge extraction is stabilizing and the parsing is stopped.

[0024] S2.2. Perform expert agent knowledge extraction, specifically including: unit text design element extraction, design element semantic alignment, design element structural standardization, design element association design technique extraction, design technique semantic alignment, design technique structural standardization, unit case image design element extraction, and unit case design element association design technique extraction. The unit text design element extraction and design element association design technique extraction adopt an iterative mechanism.

[0025] S2.3. Perform public intelligent agent knowledge extraction, wherein the thought chain guides the public intelligent agent to perform sequential tasks under the standardized system established by experts, including: identification of design elements of unit case images and extraction of design techniques associated with design elements of unit cases;

[0026] S2.4. Summarize the frequency of elements and techniques identified by the public intelligent agent with the parsing results of the expert intelligent agent. Taking practical cases as units, determine the importance attributes in the knowledge graph by statistically analyzing the total frequency of the design elements and design techniques in all expert texts and public identification cases.

[0027] As a preferred technical solution, the multimodal knowledge graph of element-method-case-image mapping is formed by coupling text modal entities and image modal entities through multidimensional relationships. The specific construction steps include:

[0028] The topological structure of the multimodal knowledge graph is established, the professional attributes and hierarchical logic of entities are defined, the relationship logic and relationship attributes between entities are defined, the entities are connected into a semantic network through multidimensional relationship edges, and the quantitative features carried by the relationships are clarified to realize the mapping of elements-techniques-cases-images. The feature weight visualization of the knowledge graph is performed. The feature weight visualization includes using the size of the nodes to reflect the frequency of the design elements or design techniques in the case data, using the color of the nodes to distinguish the design techniques under different categories of attributes, using the color of the relationship lines between entities to distinguish the association type, namely basic association relationship and co-occurrence relationship, and using the thickness and color intensity of the connection to represent the co-occurrence intensity between design techniques.

[0029] As a preferred technical solution, the knowledge enhancement prompt engineering includes trigger words, knowledge enhancement prompts, and user-defined additional descriptions; the trigger words are used to activate the LoRA model weights; the knowledge enhancement prompts provide the knowledge graph with recommended content combining high- and low-frequency design techniques; and the user-defined additional descriptions are used to input personalized design requirements.

[0030] As a preferred technical solution, the texture patching LoRA model is combined with local redrawing technology to identify the current background mask image of the area to be designed, and trigger words are called to drive the texture patching LoRA model to generate the planar texture bitmap in batches within the mask area.

[0031] As a preferred technical solution, the two-dimensional linkage simulation transforms the planar texture bitmap into a vector building outline as a base, and vertically stretches each closed building outline according to the physical height attribute obtained by the building physical height and image grayscale mapping criterion to generate a three-dimensional building volume model with spatial semantics.

[0032] As a preferred technical solution, the building physical height and image grayscale mapping criterion is a cross-dimensional data transformation mechanism, specifically including:

[0033] In the data preprocessing stage, the physical height of buildings in the traditional urban fabric map is proportionally normalized and mapped to image grayscale feature values ​​to generate a height field grayscale map, and a one-to-one correspondence benchmark between grayscale values ​​and building physical height is established.

[0034] In the two-dimensional and three-dimensional linkage simulation stage, the grayscale information in the generated planar texture bitmap is reverse-analyzed, and the grayscale information is restored to the physical height attribute of the building. Combined with the vector building outline, vertical stretching is completed to realize the mapping from two-dimensional planar texture to three-dimensional building volume.

[0035] As a preferred technical solution, the facade appearance is achieved collaboratively using depth constraints, the knowledge-enhanced cue word engineering, and the knowledge-enhanced facade design LoRA model. Specific steps include:

[0036] Extract the facade screenshots of the three-dimensional building mass model, obtain a depth control map through a depth-of-field algorithm as the depth constraint, and combine the knowledge-enhanced prompt words to output the final distinctive style design scheme.

[0037] The depth constraint is a spatial depth geometric constraint constructed based on the spatial geometric features of the three-dimensional building volume model, used to constrain the facade generation effect, and is specifically implemented using a depth control map.

[0038] The depth-of-field algorithm is an algorithm for extracting and quantizing spatial depth information from the front elevation screenshot of a 3D building mass model. It identifies the spatial relationship between the near and far views of each component and region in the 3D building mass and converts the spatial relationship into a grayscale depth-of-field control map represented by pixel brightness values.

[0039] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0040] (1) This invention introduces a thinking chain-guided multi-agent system to analyze multimodal data from two perspectives, construct a multimodal knowledge graph of elements-methods-cases-image mapping, and extract professional method tags as generation drivers to provide systematic knowledge enhancement for the generation of the final urban characteristic landscape design scheme. This solves the problems of lack of professional logic guidance, insufficient professionalism of the generated results and excessive human workload in the existing technology. It achieves the technical effect of producing highly professional and quantitative alternative schemes within a limited time, improves the overall efficiency of scheme generation, and enhances the social usability of design knowledge.

[0041] (2) Based on the principle of mapping physical height of buildings and grayscale of images, this invention establishes a progressive workflow of texture patching, block construction and facade generation. It utilizes the height physical information carried by the grayscale semantics in the two-dimensional planar texture generated by divergence to generate three-dimensional blocks. Then, it extracts the depth control map with depth information from the blocks as a geometric constraint to control the facade generation, ensuring a strict logical correspondence between the two-dimensional appearance and the three-dimensional space. It solves the problem that the system integration degree of the whole process design is extremely low and the results are difficult to transform due to the break of cross-dimensional semantics. It realizes that the generated scheme not only has a realistic visual effect, but also has the technical effect of spatial structure and construction details that conform to the logic of architectural profession.

[0042] (3) This invention introduces an automatic verification based on basic morphological indicators for the planar texture patching process. By judging whether the statistical characteristics of the generated scheme fall within the confidence interval of the specific urban characteristic texture, the scientific evaluation and elimination of the planar texture authenticity is carried out, ensuring that the scheme is unified with the diversity of design and the purity of style logic. It solves the problem of lacking an effective screening mechanism and making it difficult to verify the consistency of style from the perspective of urban morphology, ensuring the scientific nature and feasibility of the scheme, and strictly meeting the requirements of urban design guidelines in core dimensions such as construction intensity, plot ratio and spatial layout.

[0043] (4) This invention generates the planar texture of the plot through the LoRA model of texture patching and combines it with local redrawing. It realizes the divergent exploration of design in the process of planar texture patching and facade appearance expression. It solves the problem of logical convergence and rigid and single output scheme of existing rule- or instruction-driven technical solutions. It achieves the technical effect of quickly providing dozens of different possible spatial layout schemes for decision-making reference at extremely low marginal cost, thereby greatly reducing the design rework cost caused by incorrect direction and has significant economic benefits. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the process of the present invention;

[0045] Figure 2 This is a schematic diagram of the standardized texture sampling and preprocessing process;

[0046] Figure 3 This is a statistical distribution map of local textural morphology indicators;

[0047] Figure 4 A flowchart for extracting and analyzing knowledge about the unique features of a city;

[0048] Figure 5 A multimodal knowledge graph structure diagram for urban character features;

[0049] Figure 6This is a schematic diagram of the entire process of generating and filtering the texture, volume, and facade of the new Suzhou-style architecture in this embodiment using a two-dimensional linkage. Detailed Implementation

[0050] 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0051] To address the problems existing in the aforementioned prior art, this embodiment provides a full-process design method for distinctive architectural styles that combines knowledge enhancement and 2D / 3D linkage. Taking the creation of a new Suzhou-style architectural style in a waterfront building complex in Yangzhou City, Jiangsu Province as an example, this invention will be further explained.

[0052] Urban character refers to the overall visual and spatial features of a specific urban area, comprised of spatial form, architectural style, and cultural imagery. In this invention, the term focuses on the architectural expression aspect of urban character, including planar texture, three-dimensional volume, and facade appearance.

[0053] Full-process design refers to the technical pain point of the disconnect between various stages of existing generative design, including texture, volume, and facade. It is a progressive design method built through intelligent means, covering planar texture restoration, three-dimensional block construction, and facade image generation. This method starts with the logic of planar space and ensures the consistency of each design stage at the geometric and semantic levels through a two-dimensional and three-dimensional linkage mechanism, thereby achieving a more efficient solution derivation than traditional manual design.

[0054] like Figure 1 As shown, the method steps of the present invention include the following:

[0055] S1. Obtain multimodal data of the target landscape type, establish a mapping criterion between building physical height and image grayscale, and perform height information grayscale preprocessing and standardized sampling unit spatial cutting on the local texture vector map in the multimodal data based on the mapping criterion between building physical height and image grayscale. Extract morphological indicators of texture samples within the sampling unit, define the confidence interval of local texture typology indicators, and establish a feature library of local texture morphology indicators.

[0056] First, this embodiment involves multimodal raw data collection for the Neo-Soviet style. The data covers three categories: first, academic theoretical texts, including relevant academic literature, books, and dissertations; second, excellent practical case study images and accompanying design descriptions, selecting 102 images of excellent practical cases with relatively independent facade information; and third, local traditional urban fabric maps containing building physical height information, obtaining floor plans of traditional buildings around the research object that reflect local characteristics.

[0057] Secondly, the collected data is categorized and processed. For case elevation drawings, the 4X-UltraSharp or R-ESRGAN-4x algorithms are used for magnification, uniformly increasing the resolution to 1024×1024 pixels for subsequent training of the S3 mid-facade LoRA model. For local traditional urban fabric maps, the standardized processing procedure is as follows: Figure 2 As shown, this embodiment performs grayscale processing on height information, normalizing the physical height of the building proportionally and mapping it to image grayscale feature values, generating a height field grayscale map that carries the spatial height characteristics, and establishing a benchmark for the mapping between physical height and image semantics in the two-dimensional and three-dimensional linkage inference process, namely, the mapping criterion between building physical height and image grayscale. Subsequently, based on the maximum perceptible distance of 40m in urban public space, a 40m×40m square sampling unit is used to spatially segment traditional urban fabric samples, obtaining a total of 24 standardized fabric samples that carry specific spatial rhythms and enclosure relationships, and saving them in an image format that can be used for fine-tuning model training.

[0058] The "two-dimensional / three-dimensional linkage" in this invention refers to a cross-dimensional data correspondence mechanism based on the mapping criterion between building physical height and image grayscale. During model training, this mechanism converts the height information of a three-dimensional building into grayscale values ​​of a two-dimensional image, i.e., high-brightness, low-black pixel encoding, enabling the generative model to learn planar texture features with spatial height constraints. During the design and generation phase, the algorithm reverse-analyzes the grayscale data in the generated image, restoring it to vector height and performing three-dimensional stretching, thus achieving a closed-loop linkage from three-dimensional feature pixelation to two-dimensional pixel vectorization. This linkage essentially constructs a geometrically consistent chain from planar semantics to spatial entities and then to visual presentation.

[0059] Finally, a local texture morphology index feature library was established. Computational geometry algorithms were used to extract indices from texture samples within 24 sampling units. These indices included: number of buildings, average number of floors, base area, total building area, floor area ratio, building density, and open space ratio. Figure 3As shown, this embodiment uses statistical distribution analysis to define confidence intervals for indicators of specific urban characteristics, such as 99% confidence intervals. The mean and standard deviation of each morphological indicator are calculated and used as a screening benchmark for the planar texture divergence generation scheme in the subsequent S4 step. By comparing whether the indicators of the generated schemes fall within the confidence intervals of the statistical characteristics, outlier schemes are automatically identified and eliminated.

[0060] Morphological indicators refer to geometric and statistical parameters used to quantitatively describe the characteristics of urban spatial organization. In this invention, these indicators include, but are not limited to, the number of buildings, building density, floor area ratio, and average number of floors, serving as automatic verification standards after texture generation to evaluate the consistency between the generated scheme and the local traditional texture in terms of spatial layout logic.

[0061] S2. Construct a multi-agent system with dual expert and public perspectives. Utilize a thought chain to guide this system in collaborative multi-agent analysis of multimodal data. Summarize and quantify the analysis results, constructing a multimodal knowledge graph with element-method-case-image mapping that features co-occurrence quantification of design techniques and visualization of feature weights. Based on this knowledge graph, build a recommendation system for high- and low-frequency design technique combinations. The specific implementation steps are as follows:

[0062] 1. Constructing a Dual-Perspective Intelligent Agent System for Experts and the Public. This embodiment utilizes a Large Language Model (LLM) to construct a multi-agent system (MAS) with differentiated cognitive dimensions, enabling in-depth deconstruction and intuitive capture of urban landscape knowledge. The Large Language Model refers to a complex machine learning model pre-trained with massive parameters and multidimensional datasets, possessing deep semantic understanding, logical reasoning, and multimodal data parsing capabilities. In this invention, it serves as the cognitive driving engine of the multi-agent system, realizing the extraction, semantic alignment, and logical reconstruction of fragmented architectural domain knowledge.

[0063] First, an expert-perspective intelligent agent system is constructed. In this embodiment, a specific identity protocol is formulated for the large language model, positioning it as an architect and architectural style analysis expert with profound professional knowledge. This intelligent agent possesses a deep understanding of the modern inheritance of the traditional Suzhou architectural style. Its core task is to analyze the academic theoretical texts and practical case study materials obtained from S1 step by step from a professional perspective. By endowing it with a professional personality, the system is ensured to accurately extract industry-consensus design principles from complex architectural corpora.

[0064] Secondly, a public-perspective intelligent agent system is constructed. This embodiment utilizes a large language model to simulate various types of public identities, constructing 40 virtual public intelligent agents covering different ages, professional backgrounds, and aesthetic preferences. This intelligent agent system is assigned a spatial user identity based on intuitive perception, and its core task is to identify visual features in 102 excellent practical case images that align with public interest. By introducing a public perspective, this invention can supplement professional perspectives with social aesthetic preferences, thereby enhancing the social acceptability of the generated results.

[0065] Finally, the logical sequence of the dual-perspective collaborative analysis is established. The expert agent focuses on in-depth analysis of academic theories and design logic to establish a terminological benchmark for professional design elements and their techniques; while the public agent, under this benchmark, focuses on identifying interests in the sensory experiences of the built environment. Through sequential collaboration, the two types of agents aim to transform fragmented urban landscape information into a unified representation within a professional terminology system, laying the foundation for identity and logical benchmarks for the subsequent step-by-step extraction task using the CoT (Conceptual Chain of Thought).

[0066] 2. Utilizing the Chain of Thought (CoT) to guide multi-agent collaborative analysis of multimodal data, extracting design elements and techniques. The Chain of Thought refers to a prompting strategy that decomposes complex problems by guiding a large language model to generate a series of intermediate reasoning steps. In this invention, the Chain of Thought serves as a logical guiding protocol for multi-agent collaborative work, guiding agents to execute sequential reasoning step-by-step: design element extraction—semantic alignment—structured classification. Multi-agent refers to a collaborative computing architecture composed of multiple agents with independent role attributes and knowledge bases. In this invention, it includes expert agents simulating architectural professional logic and public agents simulating social aesthetic preferences, transforming fragmented information into a structured multimodal knowledge graph through dual-perspective interaction. For example... Figure 4 As shown, Figure 4 This embodiment demonstrates the entire knowledge extraction process, guided by the thought chain, from raw multimodal data to a standardized design method library. This embodiment utilizes the thought chain as a logical guiding protocol for a multi-agent system, guiding expert and public agents to complete knowledge extraction through sequential reasoning. The specific implementation steps are as follows:

[0067] First, the overall scale of the parsing task and the criteria for determining knowledge saturation are established. This embodiment processes approximately 250,000 words of academic theoretical text, including academic papers and monographs, as well as 102 excellent practical cases with illustrations and explanations. The expert agent employs an incremental iterative mechanism during extraction: using 15% of the entire corpus as the initial parsing unit, and then expanding the corpus step by step in increments of 5%. When, within two consecutive step increments (i.e., adding 10% of the corpus), the number of newly extracted design elements falls below 1% of the total accumulated elements, the system determines that knowledge extraction has stabilized and stops iterative text parsing.

[0068] Secondly, the expert agent knowledge extraction stage is executed. Knowledge extraction refers to the process of identifying and extracting entities, attributes, and relationships from unstructured or semi-structured data sources to construct a knowledge base. In this invention, it specifically refers to the technical path of identifying and extracting architectural design entities, including design elements, design techniques, and case information, as well as their interaction relationships. This embodiment breaks down the task of this stage into the following actions using a thought process:

[0069] A. Extraction of design elements from unit text: The expert agent is instructed to identify nouns related to architectural design in the text as original design elements.

[0070] B. Semantic alignment of design elements: guide the expert agent to execute the ontology alignment strategy, that is, compare the original elements identified. If multiple terms point to the same entity, such as "window", "window", and "exterior window", then map them uniformly to the standard term "window" according to the semantic center.

[0071] C. Structural standardization of design elements: The expert agent performs a structured classification of the standardized elements according to their spatial affiliation, forming a terminology system that includes the overall shape, building interface, accessories, and the surrounding environment of the building.

[0072] Semantic alignment and structural standardization of design elements include: merging synonymous or near-synonymous elements in different contexts, and structurally classifying the elements after eliminating redundancy to form a terminology system for architectural design elements with consistent categories and hierarchical divisions.

[0073] D. Extraction of Design Techniques Related to Design Elements: An expert agent analyzes the representation of elements identified in the text. For example, if the identified element is "roof," then the corresponding specific processing techniques are extracted, such as "pitched roof" or "reverse gable roof," thus establishing a parent-child mapping relationship between elements and techniques. This step aims to establish the association mapping between design elements and design techniques.

[0074] E: Semantic Alignment of Design Techniques: Guides LLM to execute multi-dimensional semantic deconstruction instructions. For simple techniques, synonym merging is performed, such as merging "white wall" into "white wall"; for complex techniques, tag-based splitting is performed, such as automatically deconstructing "white plastered wall" and mapping it into two independent tags, "white (color attribute)" and "plaster (material attribute)," to achieve atomic storage of features.

[0075] F: Structural standardization of design techniques: Based on the dimensions of technique expression, they are categorized into types such as geometric shapes, material matching, color matching, and overall intent, forming a complete professional design technique library.

[0076] Semantic alignment and structural standardization of design techniques include: merging synonymous or near-synonymous expressions of techniques, and categorizing them structurally according to the type of technique to form a standardized terminology system for design techniques covering types such as color, material, geometry, and style.

[0077] G. Extraction of design elements from unit case images: The system uses the standard terms of the above design elements as an index, and through multimodal parsing capabilities, identifies and extracts concrete design elements that can reflect the characteristics of the new Suzhou-style city from 102 practical case images, realizing the initial mapping from abstract text terms to concrete image features.

[0078] H. Extraction of design techniques for unit case design element association: The expert agent analyzes the manifestation of the elements identified in the image in specific cases.

[0079] Subsequently, the knowledge extraction phase of the public intelligent agent is executed. This embodiment guides the public intelligent agent to perform sequential tasks within a standardized system established by experts, using a thought chain, to integrate social aesthetic preferences. The specific steps are as follows:

[0080] I. Design Element Recognition in Unit Case Images: Based on the design element terminology system established by expert agents, 40 public agents with different identity attributes each identify design elements in 102 case images that conform to their specific aesthetic preferences. Specifically, the system adopts a collective consciousness integration mechanism based on majority consensus: for each case image, the frequency of elements identified by all public agents is counted, and only elements identified by more than 50%, i.e., more than 20 agents, are retained as public preference elements, and their frequency of occurrence in the image is recorded. Furthermore, the system ensures that this set of elements is precisely aligned semantically with the expert terminology system.

[0081] J. Unit Case Design Element Correlation and Design Technique Extraction: Based on the public preference elements identified above, further extract the specific design techniques associated with them in the case. This step establishes a correlation mapping between public perception intentions and professional design techniques by simulating the public's intuitive tendencies in perceiving architectural style.

[0082] Finally, multi-perspective knowledge is summarized and importance attributes are quantified. The system summarizes the frequency of elements and techniques identified by the public intelligent agent with the analysis results of the expert intelligent agent. Specifically, this embodiment determines the importance attributes of the entity in the knowledge graph subsequently built in this step by statistically analyzing the total frequency of each design element and design technique in all expert texts and public identification cases, thus constructing a standardized design technique library.

[0083] 3. Constructing a multimodal knowledge graph mapping elements, techniques, cases, and images. A multimodal knowledge graph is a structured semantic network capable of storing and associating information from multiple modalities. In this invention, the graph uses architectural design entities as nodes and the co-occurrence intensity and mapping relationships between entities as edges, serving as a knowledge interface to provide deterministic guidance on professional design techniques to the generation engine. This embodiment restructures the standardized landscape knowledge extracted and aligned through dual-view agents in the preceding steps, constructing a searchable, callable, and model-learnable multimodal knowledge graph of urban landscape characteristics. The specific implementation steps are as follows:

[0084] First, establish the topological structure and entity definitions of the multimodal knowledge graph. For example... Figure 5 As shown, this graph is formed by coupling text modal entities and image modal entities through multidimensional semantic relationships. The text modal entities include the subcategories of architectural design elements, design techniques, and case names; the image modal entities are the corresponding case images.

[0085] Secondly, define the entity's professional attributes and hierarchical logic. For example... Figure 5 As shown, this embodiment assigns detailed attribute descriptions to various entities in the atlas to define the application context of landscape knowledge:

[0086] A. Architectural Design Element Attributes: These attributes embody the hierarchical classification determined by the structural standardization process of design elements, enabling precise positioning of specific urban characteristic elements within the terminology system. Their importance is determined by the frequency of their appearance across all cases.

[0087] B. Design Technique Entity Attributes: It has an importance attribute, which is determined by the frequency of the design technique in all cases.

[0088] C. Case Name Entity Attributes: Integrates multi-dimensional project background attributes, including project function, project scale, project construction type, administrative location, and project building area, realizing a complete mapping from design techniques to specific construction context.

[0089] D. Case Image Entity Attributes: Possesses image perspective attributes, used to label the visual presentation dimensions of the image, such as outdoor human view effect diagrams.

[0090] Subsequently, the relationship logic and attributes between entities are defined. This embodiment connects heterogeneous entities into a semantic network through multi-dimensional relationship edges and clarifies the quantitative characteristics carried by the relationships:

[0091] A. Technique Type Relationship: Used to connect the subcategories of architectural design elements with the design technique entities.

[0092] B. Image Relationships: Used to connect design technique entities with concrete case image entities to achieve cross-modal alignment.

[0093] C. Inclusion Relationship: Used to connect the case name entity and the case image entity, ensuring that the image data is logically bound to the specific project background information.

[0094] D. Co-occurrence Relationship and its Attributes: This relationship exists between different design techniques and is used to describe the logic of different techniques appearing together in a given case. In particular, this relationship has a co-occurrence strength attribute, which is used to quantify the frequency of different techniques appearing together, thereby revealing high-frequency combination patterns in landscape design.

[0095] Finally, a quantitative and visual representation of the knowledge graph is performed. This embodiment uses visualization to intuitively reveal the core features and potential coupling relationships of the urban landscape knowledge system:

[0096] A. Node visualization: The size of the node reflects the frequency of a design element or technique in the case data, i.e., the importance attribute. The larger the node, the more prominent its core position in the style system. At the same time, the color of the node is used to distinguish the design techniques under different categories.

[0097] B. Relationship Visualization: Basic associations and co-occurrence relationships are distinguished by the color of the connecting lines. The thickness and color intensity of the connecting lines represent the co-occurrence intensity of design techniques, thereby revealing the combination patterns of design techniques in urban character design practice and providing a regular design reference for the subsequent generation process.

[0098] By constructing this multimodal knowledge graph, this invention transforms fragmented design experience into a structured knowledge interface, which not only achieves the traceability of design knowledge, but also provides definite logical guidance for the subsequent S3 process of using knowledge to guide model fine-tuning and constructing a regular prompt word project.

[0099] S3. Knowledge-Enhanced Generative Model Fine-tuning and Prompt Word Engineering Construction: A texture-mending LoRA model is trained based on a standardized two-dimensional local texture map obtained from standardized sampling unit spatial segmentation. A knowledge-enhanced facade design LoRA model is trained based on data from a multimodal knowledge graph. A knowledge-enhanced prompt word engineering is constructed based on a combination of high- and low-frequency design techniques. Knowledge enhancement refers to the use of externally structured prior knowledge to intervene in the generative model, thereby improving its logical accuracy and performance stability in a specific professional field. In this invention, this method specifically refers to injecting architectural design logic and stylistic features into the entire generation process. This method is achieved through a dual-path collaborative approach of parameter-level LoRA fine-tuning and logic-level guided prompt word engineering. Prompt word engineering refers to the technique of guiding the model to output specific content through explicit language instructions. In this invention, high- and low-frequency design techniques retrieved from the knowledge graph are structurally combined to form prompt instructions with professional logical guidance, thus controlling the generation results. The urban 2D texture features acquired in S1 and the structured knowledge constructed in S2 are transformed into the core driving force of generative AI. Through model fine-tuning and instruction engineering, a precise transformation from knowledge storage to design generation is achieved. The specific implementation steps are as follows:

[0100] 1. Training and Fine-tuning of a Dual LoRA Model Based on a Multimodal Dataset. This embodiment utilizes the semantic and spatial constraints provided by a multimodal knowledge graph to construct dual LoRA models for urban 2D texture and building facades respectively, achieving precise and controlled generation of urban landscape features. Low-Rank Adaptation (LoRA) is an algorithm module that achieves efficient model transfer by decomposing the pre-trained weight matrix into a low-rank matrix product without changing the full parameters of the original model. In this invention, this module serves as a parameter-level carrier for knowledge enhancement, used to inject landscape features into the model's latent space, transforming random generation logic into a design expression that better conforms to professional standards.

[0101] A. Training of the LoRA model for texture patching. The dataset consists of standardized sampling unit images obtained from S1, totaling 24 ground planar texture samples. Each image in this dataset is matched with a specific trigger word as a label; in this embodiment, it is "New_Su_School_planar texture". This model is used to activate the specific urban characteristic spatial rhythm and enclosure relationship learned by the model during the generation stage, realizing the continuation and patching of urban texture at the street level. In specific training, the system selects F.1_dev-fp8 as the basic pre-trained model. Its training hyperparameters can be dynamically adjusted according to the corpus size and hardware environment. In this embodiment, the preferred settings are: 15 repetitions, 20 training rounds, batch size of 1, and the AdamW8bit optimizer is used with a learning rate of 1×10-4, a Network Alpha of 32, and a noise offset of 0.1. This configuration can effectively guide the model to accurately reproduce the spatial organization logic of the street at a macro scale.

[0102] Planar texture patching refers to the process of completing a scheme through divergent deduction of the spatial organization of buildings within a specific design scope. In this invention, planar texture uses semantically encoded images of building height as a carrier, that is, quantitatively encoding the physical height of buildings through the grayscale values ​​of pixels to characterize the spatial morphological features of building clusters, such as density, outline, and height distribution. Trigger words are semantic identifiers used in the prompt word project to activate specific LoRA models or specific design logic patterns. In this invention, trigger words serve as a semantic interface connecting knowledge-guided prompts and model fine-tuning parameters, used to precisely and directionally activate the urban characteristic features injected into the model during the diffusion model inference stage, including architectural spatial texture and facade design logic. By including specific trigger words in the prompt words, such as "New_Su_school" or "New_Su_School_planar texture" in this embodiment, this invention can efficiently map external user design intentions with internal professional landscape knowledge, thereby achieving a professional transformation from general image generation to the shaping of specific urban architectural characteristics.

[0103] B. LoRA Model Training for Facade Design. The dataset is based on the cross-modal mapping relationship of the multimodal knowledge graph in S2. The system extracts 102 image samples labeled with aesthetic consensus from both expert and public perspectives, based on the link logic between associated images and design elements / techniques in the knowledge graph. By combining the trigger word "New_Su_school" with the extracted design technique labels, a facade fine-tuning dataset with aesthetic consensus labels from both professional and public perspectives is constructed, ensuring that the model establishes an accurate style reference during the generation process. In specific training, the system uses the same training hyperparameter configuration as the texture patching model mentioned above, i.e., performing weight optimization under the same repetition count, training epochs, and learning rate settings. This configuration is sufficient to ensure that the model weights converge accurately with the trend of the loss function, thereby establishing a style reference corresponding to the standardized design techniques in the knowledge graph within the generation engine, completing an effective transfer from structured knowledge to model-generated weights.

[0104] 2. Construction of Knowledge-Enhanced Prompt Lexicon and Recommendation Logic. This embodiment constructs a prompt lexicon driven by a knowledge graph, guiding the model to achieve knowledge enhancement and result control without changing parameters through explicit language instructions.

[0105] The prompt vocabulary consists of three parts: trigger words, knowledge enhancement prompts, and user-defined additional descriptions. Trigger words activate specific weights in the fine-tuned LoRA model, enabling knowledge enhancement generation during the invocation phase. Knowledge enhancement prompts anchor features based on professional design knowledge provided by the knowledge graph, transforming standardized design techniques from the graph into professional logical instructions to guide model generation. User-defined additional descriptions allow users to input descriptive words based on the personalized needs of specific projects, such as environment, time, or detailed requirements, to enhance the diversity and controllability of the generated results.

[0106] The prompt word library features a recommendation mechanism based on knowledge graph attributes, supporting flexible combinations of design techniques. Logical instructions are implemented by retrieving design elements with different importance attributes from the graph. The system recommends high-frequency techniques to highlight the overall macro-level features, ensuring consistency in the generated results across a systematic dimension. Simultaneously, it supports the introduction of low-frequency design techniques to reflect differentiated characteristics within a macro-context, meeting the innovative needs of specific scenarios. Finally, the system performs path retrieval in the graph based on the user's input design intent, automatically assembling the above three parts into structured prompt words.

[0107] S4. Using a trained LoRA model for texture patching, and combining local redrawing with divergent generation of planar textures, a two-dimensional planar texture bitmap is obtained. The planar texture bitmap is converted into a vector building outline. Simultaneously, based on the mapping criterion between building physical height and image grayscale, the grayscale information of the planar texture bitmap is converted into the building physical height attribute of each closed vector building outline. Based on the vector building outline and building physical height attribute, morphological base data is formed. According to the morphological base data, full-dimensional morphological indicators are calculated. The full-dimensional morphological indicators are compared with the confidence interval of the local texture typology indicators in real time, and the schemes with indicator distributions that meet the constraints are selected. Using the vector building outline and the mapping criterion between building physical height and image grayscale, two-dimensional and three-dimensional linked block stretching is performed to obtain a three-dimensional building volume model. Combined with knowledge-enhanced prompt word engineering, the facade style is generated collaboratively using depth constraints and a trained knowledge-enhanced facade design LoRA model, and the final urban characteristic style design scheme is output.

[0108] This step, serving as the execution terminal of the technical solution of this invention, transforms the structured knowledge and fine-tuning model generated in the preceding steps into a feasible design solution through multi-stage linkage and screening. For example... Figure 6 As shown, this embodiment performs the following steps on a specific plot of land in Yangzhou City where a new building is to be constructed:

[0109] 1. Divergent Generation of Planar Texture. This embodiment first utilizes a finely tuned LoRA texture patching model combined with inpainting technology to achieve divergent generation of the planar texture of the site. First, for the site to be designed, a 40m square area with the site as its core is identified as the patching unit, and the existing base map mask image of the area to be designed within this unit is identified. By calling the corresponding trigger word "New_Su_School_planar texture", the generation model is driven to generate multiple sets of planar texture bitmaps with specific spatial rhythms and enclosure relationships within the mask area. This step, through the scale constraint of the 40m unit, ensures that the generated texture possesses a spatial modulus that is compatible with the surrounding traditional landscape in the initial stage.

[0110] Local redrawing refers to image generation techniques that complete or reconstruct content for specific masked regions in an image. In this invention, this technique is used in the planar texture patching stage, that is, while preserving the existing texture around the area to be designed, the texture patching LoRA model is used to perform divergent morphological deduction within the redrawing area.

[0111] 2. Two- and three-dimensional linked vectorization conversion and morphological index calculation based on building physical height and image grayscale mapping criteria. The generated two-dimensional texture bitmap is then subjected to two- and three-dimensional linked vectorization conversion. The bitmap is converted into a vector building outline using an image contour extraction algorithm, and according to the building physical height and image grayscale mapping criteria established in S1, the pixel grayscale information in the image is converted into building physical height attributes, thereby generating vector polygon data with three-dimensional attribute information. Based on the above vector data, morphological index calculation includes full-dimensional morphological indicators such as the number of buildings, average number of floors, base area, total building area, plot ratio, building density, and open space ratio.

[0112] 3. Automatic verification and screening based on morphological indicators. The system compares the calculated morphological indicators with the confidence intervals of the local traditional texture indicators determined in S1 in real time. Outlier schemes that exceed the confidence intervals for key indicators such as plot ratio and building density will be automatically eliminated by the system, retaining only qualified schemes whose indicator distribution characteristics are highly similar to those of Yangzhou's local traditional samples. This ensures the scientific nature of the design scheme and the rigorous inheritance of regional features from the perspective of spatial organization.

[0113] 4. Mass Extrusion Model Based on Vector Profile and Height Attributes. The system uses a qualified vector building profile as a base and performs vertical extrusion based on the mapped physical height attribute to generate a 3D building mass model with spatial semantics. This step completes the logical transformation from 2D planar data to a 3D spatial model. Before subsequent generation, this mass model allows for optional manual optimization of geometric details such as roof forms (e.g., pitched roofs, cantilevered eaves) according to actual design requirements.

[0114] 5. Depth-Constrained Facade Generation. The controllable generation of the facade style is achieved collaboratively using depth constraints and a fine-tuned LoRA model of the facade design. A screenshot of the 3D building mass's front facade is extracted, and a depth map is obtained through a depth-of-field algorithm. This map serves as the spatial geometric constraint for the ControlNet. Subsequently, combined with a prompt word project recommended by S2 knowledge graph path retrieval, including trigger words, knowledge enhancement hints, and user-defined descriptions, accurate rendering of the facade details is achieved while maintaining spatial logical consistency. The final generated scheme visually conforms to the professional aesthetics of the New Suzhou style, and all construction indicators meet the required standards.

[0115] In this invention, the depth control map refers to a grayscale image representing the distance between the surface of an object and the viewpoint in a scene, with its pixel brightness values ​​showing a specific mapping relationship with spatial depth information. In this invention, this map is generated from a 3D building block model using a depth extraction algorithm and input as a strong geometric constraint into the generative control network. This constraint precisely controls the perspective structure, scale ratio, and volumetric occlusion relationships of the generated image, ensuring that the generated facade image is strictly aligned topologically with the underlying 3D spatial logic. The control network is a neural network architecture that achieves geometrically controlled image processing by introducing additional spatial constraints. In this invention, it serves as a geometric guidance module for 2D and 3D linkage, ensuring a strict correspondence between the generated image and the 3D spatial logic, thereby effectively eliminating spatial illusions.

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for designing distinctive landscapes across the entire process, characterized by knowledge enhancement and 2D / 3D linkage, wherein... The method includes: S1. Acquire multimodal data of the target landscape type, establish a mapping criterion between building physical height and image grayscale, and based on the mapping criterion between building physical height and image grayscale, perform grayscale preprocessing of height information and standardized sampling unit spatial cutting on the local texture map in the multimodal data, extract morphological indicators of texture samples in the sampling unit, define the confidence interval of local texture typology indicators, and establish a feature library of local texture morphology indicators; The multimodal data includes academic theoretical texts, practical case studies with images and texts, and local texture maps; the local texture maps are vector maps of the architectural textures within the target landscape type of urban area that carry local characteristics; the morphological indicators are a set of geometric and statistical parameters that quantitatively describe the spatial organization characteristics of the city; the indicators in the local texture morphological indicator feature library include test indicators for the number of buildings, average number of floors, base area, total building area, plot ratio, building density, and open space ratio; S2. Construct a multi-agent system with dual perspectives of experts and the public, and use a thought chain to guide the multi-agent system to perform multi-agent collaborative analysis of the multimodal data. Then, summarize and quantify the analysis results, construct a multimodal knowledge graph with element-method-case-image mapping that has co-occurrence quantification of design techniques and visualization of feature weights, and construct a recommendation of high- and low-frequency design technique combinations based on the knowledge graph. S3. Train a texture patching LoRA model based on the standardized two-dimensional local texture map obtained by spatial cutting of the standardized sampling unit; train a knowledge-enhanced facade design LoRA model based on the data in the multimodal knowledge graph; and construct a knowledge-enhanced prompt word project based on the combination of high- and low-frequency design techniques. S4. Using the trained texture patching LoRA model, combined with local redrawing divergent generation of planar texture, a two-dimensional planar texture bitmap is obtained; the planar texture bitmap is converted into a vector building outline, and according to the building physical height and image grayscale mapping criterion, the grayscale information of the planar texture bitmap is converted into the building physical height attribute of each closed vector building outline. Based on the vector building outline and building physical height attribute, morphological base data is formed, and full-dimensional morphological indicators are calculated according to the morphological base data; the full-dimensional morphological indicators are compared with the confidence interval of the local texture typology indicators in real time, and the schemes with indicator distribution that meet the constraints are selected. Using the vector building outline and the building physical height and image grayscale mapping criterion, two-dimensional and three-dimensional linked block stretching is performed to obtain a three-dimensional building volume model. Combined with the knowledge-enhanced prompt word engineering, the facade style is generated in collaboration with the depth constraint and the trained knowledge-enhanced facade design LoRA model, and the final distinctive style design scheme is output.

2. The method for designing a distinctive landscape through knowledge enhancement and two-dimensional / three-dimensional linkage as described in claim 1, characterized in that, The specific steps for constructing the expert-public dual-perspective multi-agent system using a large language model include: Construct an expert-perspective intelligent agent system, which possesses architectural expertise, and systematically analyzes the multimodal data to extract professional design patterns. A public-perspective intelligent agent system is constructed. Based on the assigned differentiated identity profile, including age, occupation and aesthetic preferences, the public-perspective intelligent agent system simulates the visual aesthetics and interest tendencies of different social groups and identifies visual features in multimodal data that match the interests of the corresponding groups.

3. The method for designing a distinctive landscape through knowledge enhancement and two-dimensional / three-dimensional linkage as described in claim 1, characterized in that, The specific steps for using thought chains to guide the collaborative analysis of multi-agent systems from both expert and public perspectives in the multi-agent system are as follows: S2.

1. Establish the scale of the parsing task and the criteria for knowledge saturation. When the amount of new knowledge extracted is lower than the preset threshold, it is determined that the knowledge extraction is stabilizing and the parsing is stopped. S2.

2. Perform expert agent knowledge extraction, specifically including: unit text design element extraction, design element semantic alignment, design element structural standardization, design element association design technique extraction, design technique semantic alignment, design technique structural standardization, unit case image design element extraction, and unit case design element association design technique extraction. The unit text design element extraction and design element association design technique extraction adopt an iterative mechanism. S2.

3. Perform public intelligent agent knowledge extraction, wherein the thought chain guides the public intelligent agent to perform sequential tasks under the standardized system established by experts, including: identification of design elements of unit case images and extraction of design techniques associated with design elements of unit cases; S2.

4. Summarize the frequency of elements and techniques identified by the public intelligent agent with the parsing results of the expert intelligent agent. Taking practical cases as units, determine the importance attributes in the knowledge graph by statistically analyzing the total frequency of the design elements and design techniques in all expert texts and public identification cases.

4. The method for designing a distinctive landscape through knowledge enhancement and two-dimensional / three-dimensional linkage as described in claim 1, characterized in that, The multimodal knowledge graph of element-method-case-image mapping is formed by coupling text modal entities and image modal entities through multidimensional relationships. The specific construction steps include: The topological structure of the multimodal knowledge graph is established, the professional attributes and hierarchical logic of entities are defined, the relationship logic and relationship attributes between entities are defined, the entities are connected into a semantic network through multidimensional relationship edges, and the quantitative features carried by the relationships are clarified to realize the mapping of elements-techniques-cases-images. The feature weight visualization of the knowledge graph is performed. The feature weight visualization includes using the size of the nodes to reflect the frequency of the design elements or design techniques in the case data, using the color of the nodes to distinguish the design techniques under different categories of attributes, using the color of the relationship lines between entities to distinguish the association type, namely basic association relationship and co-occurrence relationship, and using the thickness and color intensity of the connection to represent the co-occurrence intensity between design techniques.

5. The method for designing a distinctive landscape through knowledge enhancement and two-dimensional / three-dimensional linkage as described in claim 1, characterized in that, The knowledge enhancement prompt word engineering includes trigger words, knowledge enhancement prompts, and user-defined additional descriptions; the trigger words are used to activate LoRA model weights; the knowledge enhancement prompts provide the knowledge graph with a combination of high- and low-frequency design techniques for recommended content; The user-defined additional description is used to input personalized design requirements.

6. The method for designing a distinctive landscape through knowledge enhancement and two-dimensional / three-dimensional linkage as described in claim 1, characterized in that, By using the texture patching LoRA model combined with local redrawing technology, the current background mask image of the area to be designed is identified, and the trigger word is called to drive the texture patching LoRA model to generate the planar texture bitmap in batches within the mask area.

7. The method for designing a distinctive landscape through knowledge enhancement and two-dimensional / three-dimensional linkage as described in claim 1, characterized in that, The two-dimensional linkage block stretching and derivation transforms the planar texture bitmap into a vector building outline as a base. Based on the physical height attribute obtained by the building physical height and image grayscale mapping criterion, each closed building outline is vertically stretched to generate a three-dimensional building volume model with spatial semantics.

8. The method for designing a distinctive landscape through knowledge enhancement and two-dimensional / three-dimensional linkage as described in claim 1, characterized in that, The building physical height and image grayscale mapping criterion is a cross-dimensional data transformation mechanism, specifically including: In the data preprocessing stage, the physical height of buildings in the ground texture map is proportionally normalized and mapped to image grayscale feature values ​​to generate a height field grayscale map, and a one-to-one correspondence benchmark between grayscale values ​​and building physical height is established. In the two-dimensional and three-dimensional linkage simulation stage, the grayscale information in the generated planar texture bitmap is reverse-analyzed, and the grayscale information is restored to the physical height attribute of the building. Combined with the vector building outline, vertical stretching is completed to realize the mapping from two-dimensional planar texture to three-dimensional building volume.

9. The method for designing a distinctive landscape through knowledge enhancement and two-dimensional / three-dimensional linkage as described in claim 1, characterized in that, The facade appearance is achieved collaboratively using depth constraints, the knowledge-enhanced cue word engineering, and the knowledge-enhanced facade design LoRA model. Specific steps include: Extract the facade screenshots of the three-dimensional building mass model, obtain a depth control map through a depth-of-field algorithm as the depth constraint, and combine the knowledge-enhanced prompt words to output the final distinctive style design scheme. The depth constraint is a spatial depth geometric constraint constructed based on the spatial geometric features of the three-dimensional building volume model, used to constrain the facade generation effect, and is specifically implemented using a depth control map. The depth-of-field algorithm is an algorithm for extracting and quantizing spatial depth information from the front elevation screenshot of a 3D building mass model. It identifies the spatial relationship between the near and far views of each component and region in the 3D building mass and converts the spatial relationship into a grayscale depth-of-field control map represented by pixel brightness values.

Images