Garden plane design method and system based on multi-dimensional constraint diffusion generation model

Abstract

The invention discloses a garden plane design method based on a multi-dimensional constraint diffusion generation model. The method comprises the following steps: constructing the multi-dimensional constraint diffusion generation model, and training based on the multi-dimensional constraint diffusion generation model to obtain a first model, a second model and a third model; in the preliminary design stage, garden design site data is input into the first model, and a plurality of different multi-dimensional preliminary design schemes are obtained; in a scheme modification stage, inputting the guidance condition and data of the remaining dimensions in the multi-dimensional preliminary design scheme into a second model to obtain a plurality of different multi-maintenance modification design schemes; and in the general graphic design stage, inputting the selected guide conditions of multiple dimensions and the site real scene data in the selected multi-maintenance-modification design scheme into a third model to obtain a plurality of different general graphic design schemes, and finally selecting out the general graphic design scheme meeting the requirements. According to the method, a designer can be helped to solve the encountered design problems more purposefully, so that the workload is reduced, and the efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence design, and in particular to a landscape plan design method and system based on a multi-dimensional constrained diffusion generation model. BACKGROUND

[0002] In recent years, artificial intelligence technology has developed rapidly, and generative models have made a great contribution to the design field with their strong creativity. Under the dual impetus of technological breakthroughs and policy guidance, the application of generative design is beginning to show its potential. The various design functions generated automatically are already quite rich, but the current generative algorithm still mainly relies on visual content creation algorithms represented by image generation tools.

[0003] The diffusion model is the most commonly used model in current design. Its main principle is a deep generative model that generates data by gradually denoising. In simple terms, it learns the process of gradually restoring a real image from pure noise, and through a U-shaped neural network (U-net), it gradually approaches the optimal solution through repeated iterations. However, according to the algorithm logic of the existing diffusion model, it cannot perfectly realize the automatic design of landscape plan scheme drawings by imitating the "three-stage" thinking of designers. The most significant defect comes from the fact that the existing diffusion model cannot learn multi-dimensional data. In the design process, designers first analyze the site design range, surrounding environment, and surrounding traffic flow; secondly, they analyze the site status: multi-dimensional constraints such as reserved vegetation, internal roads, water systems, or other existing natural resources; and finally, they combine the existing conditions to divide the functional area, construct the road network structure, determine the vegetation planting location, and deepen the node details. Obviously, the existing diffusion model can only input and output RGB three channels, and cannot achieve the above functions.

[0004] Although the existing algorithm performs outstandingly in generating images and renderings, it is difficult to complete the high-level design tasks required by designers in the actual design process, such as site understanding, concept thinking, and functional layout. This limitation leads to significant challenges for current landscape artificial intelligence design technology in creative generation, functional analysis, and layout optimization. SUMMARY

[0005] The present application is to solve the problems of the prior art by building a multi-dimensional constrained diffusion generation model (MCDM) corresponding to the three-stage framework of "preliminary design - scheme modification - general plan design", which realizes the full-process automation model of landscape plan design imitating the designer's thinking, saving time and labor costs in landscape plan design.

[0006] Technical solution: A garden plan design method based on a multi-dimensional constraint diffusion generation model, comprising the following steps:

[0007] (1) Construct a multi-dimensional constraint diffusion generation model to realize model training and design scheme process of multi-dimensional data; the design scheme process includes a preliminary design stage, a scheme modification stage and a total plan design stage; the multi-dimensional constraint diffusion generation model trains a first model in the preliminary design stage, a second model in the scheme modification stage and a third model in the total plan design stage;

[0008] (2) In the preliminary design stage, input the garden design site data into the first model to obtain a plurality of different multi-dimensional preliminary design schemes;

[0009] (3) In the scheme modification stage, introduce a conditional guided generation mechanism to take the data of a certain dimension in the selected multi-dimensional preliminary design scheme as a constraint condition, modify the constraint condition as needed and take it as a guide condition, input the guide condition and the data of the remaining dimensions in the multi-dimensional preliminary design scheme into the second model to obtain a plurality of different multi-dimensional modified design schemes;

[0010] (4) In the total plan design stage, use the conditional guided generation mechanism to take the data of multiple dimensions in the selected multi-dimensional modified design scheme generated in the scheme modification stage as optional guide conditions, input the guide conditions of the selected multiple dimensions and the site real scene data in the selected multi-dimensional modified design scheme into the third model to obtain a plurality of different total plan design schemes, and finally select a total plan design scheme that meets the requirements. In the third model, the site real scene data only provides surrounding environment information.

[0011] Further, the step (1) comprises:

[0012] (1.1) In the model training process and the design scheme process, the input multi-dimensional data formula is expressed as:

[0013] y0=[RGB,Grey,Class,Draw]∈R C×H×W

[0014] Wherein, the input multi-dimensional data y0 is tensorized and merged into a multi-channel conditional tensor R, and the shape of R is channel number C×height H×width W; RGB represents the site real scene, Grey represents the site traffic flow line, Class represents the present situation functional partition, and Draw represents the site real scene hand-drawing;

[0015] Noise predicted by the model at time step t The expression is:

[0016]

[0017] Wherein, fθ is the prediction network of the trained model, is the element-wise multiplication, M is the mask tensor, x cond is the multi-dimensional conditional input, γ t is the noise scheduling coefficient, ∈ is the standard Gaussian noise, is the noisy tensor data;

[0018] (1.2) In the model training process, the loss function is designed as The calculation formula is as follows:

[0019]

[0020] wherein, is the expected value, represents the probability weighted average of the random variable with its subscript, t is the time step, y0 is the input multi-dimensional data, ∈ k is the input real noise, k is the input of each multi-dimensional data, λ k is the proportion weight of each channel tensor, if the input is a guide condition, the corresponding proportion weight λ k = 0, is the predicted noise obtained by the model after t time steps of reasoning, ∈ k is the initial input real noise, p is a positive integer.

[0021] Further, the first model generates a random scheme in the preliminary design stage, the second model generates a scheme combined with artificial intervention in the scheme modification stage, and the third model generates a refined scheme in the overall plan design stage; the refined scheme is a scheme that can reflect a specific style.

[0022] Further, the multi-dimensional data used in the model training process includes site design range, site real scene, site traffic flow line, present situation functional partition, and site real scene hand-drawn map; in the design scheme process, the garden design site data includes site design range, site real scene, site traffic flow line, and present situation functional partition.

[0023] Further, in the process of inputting multi-dimensional data to the multi-dimensional constraint diffusion generation model, the redrawing design area is delineated through the mask mechanism, the multi-dimensional constraint diffusion generation model is input to the forward diffusion and noise adding mechanism, and after the noise adding processing, the multi-dimensional constraint diffusion generation model is input again; the noise adding mechanism is to gradually add Gaussian noise to the target area with the passage of time step t until it is indistinguishable from pure Gaussian noise.

[0024] Further, the conditional guide generation mechanism refers to taking the data of one or more dimensions in the multi-dimensional design scheme obtained in the design scheme process as a guide condition to guide the output of the next stage model; the multi-dimensional design scheme includes multi-dimensional preliminary design scheme and multi-dimensional modified design scheme.

[0025] The garden plane design system based on the multi-dimensional constraint diffusion generation model comprises a multi-dimensional constraint diffusion generation model, an information input module and an information analysis output module.

[0026] The multi-dimensional constraint diffusion generation model realizes model training and scheme design processes of multi-dimensional data; the design scheme process comprises a preliminary design stage, a scheme modification stage and a total plane design stage; the multi-dimensional constraint diffusion generation model trains to obtain a first model in the preliminary design stage, trains to obtain a second model in the scheme modification stage and trains to obtain a third model in the total plane design stage.

[0027] The information input module can pre-process and input obtained garden design site data.

[0028] In the preliminary design stage, the information analysis module inputs the garden design site data into the first model to obtain a plurality of different multi-dimensional preliminary design schemes; in the scheme modification stage, a condition guiding generation mechanism is introduced to take the data of a certain dimension in the selected multi-dimensional preliminary design scheme as a constraint condition, modifies the constraint condition as a guiding condition on demand, inputs the guiding condition and the data of the remaining dimensions in the multi-dimensional preliminary design scheme into the second model to obtain a plurality of different multi-dimensional modification design schemes; in the total plane design stage, the condition guiding generation mechanism is utilized to take the data of a plurality of dimensions in the selected multi-dimensional modification design scheme generated in the scheme modification stage as optional guiding conditions, inputs the guiding conditions of the selected plurality of dimensions and the site real scene data in the selected multi-dimensional modification design scheme into the third model to obtain a plurality of different total plane design schemes, and finally selects a total plane design scheme meeting the requirements. In the third model, the site real scene data only provides surrounding environment information.

[0029] Beneficial effects: The multi-dimensional constraint diffusion generation model MCDM designed in the application realizes the full-process automation of landscape garden plane design by simulating the designer's thinking, and more finely meets a plurality of different drawing requirements. The multi-dimensional constraint diffusion generation model MCDM has good performances in the aspects of multi-dimensional consistency, site perception, drawing diversity and controllable constraint in the landscape garden design. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 It is a schematic diagram of the multi-dimensional constraint diffusion generation model MCDM framework;

[0031] Figure 2 It is a schematic diagram of the MCDM framework in the preliminary design stage;

[0032] Figure 3 It is a schematic diagram of the MCDM framework in the scheme modification stage;

[0033] Figure 4The MCDM is used for the framework sketch of the general plane design stage.

[0034] Figure 5 The site status data graph is designed for the embodiment of the application.

[0035] Figure 6 The design effect graph is generated for the MCDM in the preliminary design stage of the embodiment of the application.

[0036] Figure 7 The design effect graph is generated for the MCDM in the scheme modification stage of the embodiment of the application.

[0037] Figure 8 The design effect graph is generated for the MCDM in the general plane design stage of the embodiment of the application.

[0038] Figure 9 The drawing example graph of the different picture generation model in the landscape garden design. DETAILED DESCRIPTION

[0039] The technical scheme of the application is further described below with reference to the drawings.

[0040] Embodiment 1

[0041] The garden plane design method based on the multi-dimensional constraint diffusion generation model comprises the following steps:

[0042] (1) Constructing a multi-dimensional constraint diffusion generation model to realize model training and design scheme process of multi-dimensional data; the design scheme process comprises a preliminary design stage, a scheme modification stage and a general plane design stage; the multi-dimensional constraint diffusion generation model trains to obtain a first model in the preliminary design stage, trains to obtain a second model in the scheme modification stage and trains to obtain a third model in the general plane design stage; the first model generates a random scheme in the preliminary design stage, the second model generates a scheme combined with artificial intervention in the scheme modification stage, and the third model generates a refined scheme in the general plane design stage; the refined scheme is a scheme capable of reflecting a specific style.

[0043] The biggest difference between the multi-dimensional constraint diffusion generation model MCDM and traditional various non-professional generation models is that the model can imitate the design idea of a designer, trains the model by using multi-dimensional data such as a site design range (Mask), a site real scene (RGB), a site traffic flow line (Grey) and a status function partition (Class), and then transmits design achievements among different models to realize deep simulation of the "three-stage" thinking process of a human designer and finally outputs a high-quality design scheme with innovation and feasibility. Figure 1 The MCDM is used for the framework sketch of the general plane design stage.

[0044] (2) In the preliminary design stage, the garden design site data is input into the first model to obtain a plurality of different multi-dimensional preliminary design schemes;

[0045] The preliminary design stage is an analysis and design process in which a designer performs brainstorming according to design site data collected in the early stage, to propose a preliminary scheme design concept. The MCDM proposed in the present application can generate a plurality of different preliminary schemes according to the early-stage investigation and analysis content input by the designer, and further provide the designer with preliminary design conception inspiration.

[0046] The input content of the MCDM in this stage is the site design range (Mask), the site real scene (RGB), the site traffic flow line (Grey), and the present situation function partition (Class), and the output content is the multi-dimensional preliminary design scheme of the site, the surrounding road, and the function partition of the site. The model converts the park real scene image as the site surrounding environment information and the present situation vegetation distribution information, the present situation function partition as the site present situation partition information, and the existing road network and other different types of data into a multi-dimensional feature tensor, processes the multi-dimensional feature tensor through the mask mechanism and the forward diffusion and noise adding mechanism, inputs the processed multi-dimensional feature tensor into the model, uses the multi-dimensional information learned in the repeated iteration training to realize internal association, completes the training, and can output a plurality of different plane layout real scenes and corresponding multi-dimensional scheme data (such as Figure 2 ) according to specific input data.

[0047] (3) In the scheme modification stage, a condition guided generation mechanism is introduced to take the data of a certain dimension in the selected multi-dimensional preliminary design scheme as a constraint condition, modify the constraint condition as needed to serve as a guide condition, input the guide condition and the data of the remaining dimensions in the multi-dimensional preliminary design scheme into the second model to obtain a plurality of different multi-dimensional modified design schemes;

[0048] The main task of the scheme modification stage is to optimize and adjust the preliminary design results based on external constraint conditions. In actual projects, the designer needs to report the multi-dimensional preliminary design scheme to the construction unit, and then modify the design according to the opinions of the construction unit.

[0049] In order to facilitate the designer to modify the original preliminary scheme, the MCDM proposed in the present application introduces a condition guided generation mechanism in this stage, and the designer can input the modified function partition design or road network and other conditions as constraint conditions to guide the generated multi-dimensional modified design scheme. Specifically, the selected multi-dimensional preliminary design scheme output in the preliminary design stage is input, and the design area that needs to be modified is processed through the mask mechanism and the forward diffusion mechanism, and a function partition guide image modified by a person is added to generate a plurality of different multi-dimensional modified design schemes, which will effectively improve the landing performance and demand matching degree of the generated scheme (such as Figure 3 ).

[0050] (4) During the site layout design phase, a condition-guided generation mechanism is used to take data from multiple dimensions of the selected multi-dimensional modified design schemes generated during the scheme modification phase as optional guiding conditions. The selected guiding conditions and the site scene data from the selected multi-dimensional modified design schemes are input into the third model to obtain multiple different site layout design schemes. Finally, the site layout design scheme that meets the requirements is selected. In the third model, the site scene data only provides information about the surrounding environment.

[0051] The main task of the master plan design phase is to conduct a more detailed design expression and visual optimization of the existing scheme, such as road network sorting and node refinement, so as to further improve the aesthetics and readability of the design scheme.

[0052] To achieve this stage, this invention adds hand-drawn landscape design plans (RGB) as a training dataset to the MCDM training. Because hand-drawn plans have a clearer road network layout and node details compared to actual park images, the model can generate more refined planar design details (such as...) under the free guidance of design requirements (whether to add road network guidance, whether to add functional zoning guidance). Figure 4 ).

[0053] Furthermore, step (1) of constructing the multidimensional constrained diffusion generation model includes:

[0054] (1.1) In the model training and design process, the multidimensional constrained diffusion generation model adds dimensions to the input and output data based on the single-dimensional image data input and output of the traditional diffusion model. The formula for the multidimensional input data is expressed as follows:

[0055]

[0056] The input multi-dimensional data y0 is tensorized and merged into a multi-channel conditional tensor R. The shape of R is the number of channels C × height H × width W. In this formula, the number of channels C = 3 + 1 + K + 3 = (7 + K)ch, that is, y0 is a tensor with a shape of (7 + K) × H × W. The value of K depends on the number of categories in the classified image. RGB represents the actual scene of the site (a 3-channel RGB image), Grey represents the traffic flow lines of the site (a single-channel grayscale image), Class represents the current functional zoning (a multi-channel TIF image), and Draw represents a hand-drawn sketch of the actual scene of the site (a 3-channel RGB image).

[0057] Noise predicted by the model at time step t The expression is:

[0058]

[0059] Among them, fθ is the trained model prediction network (i.e. U-net network), is the element-wise multiplication, M is the mask tensor (which has the same shape as tensor y0), x cond is the multi-dimensional conditional input, γ t is the noise scheduling coefficient, ∈ is the standard Gaussian noise, is the noisy tensor data.

[0060] (1.2) In the model training process, the loss function is designed as The calculation formula is as follows:

[0061]

[0062] wherein, is the expected value, represents the probability weighted average of the random variable with its subscript, t is the time step, y0 is the input multi-dimensional data, ∈ k is the input real noise, k is the input of each multi-dimensional data, λ k is the proportion weight of each channel tensor, if the input is the guide condition, the corresponding proportion weight λ k = 0, is the predicted noise obtained by the model after t time steps of reasoning, ∈ k is the initial input real noise, p is a positive integer. When p = 2, i.e. using L2 norm, the loss is calculated using mean square error (MSE) loss. Substitute the value of multi-dimensional data k into the loss function The expansion formula is as follows:

[0063]

[0064] Further, the multi-dimensional data used in the model training process includes the site design range (Mask), the site real scene (RGB), the site traffic flow line (Grey), the present situation functional partition (Class), and the site real scene hand-drawing (Draw); in the design scheme process, the garden design site data includes the site design range (Mask), the site real scene (RGB), the site traffic flow line (Grey), and the present situation functional partition (Class).

[0065] Further, in the process of inputting multi-dimensional data to the multi-dimensional constraint diffusion generation model, the design area is demarcated by the mask mechanism (i.e. white for the redrawn area and black for the non-redrawn area), and the multi-dimensional constraint diffusion generation model is inputted to the forward diffusion and noise adding mechanism after the noise processing; the noise adding mechanism is to gradually add Gaussian noise to the target area with the passage of time step t until it is indistinguishable from pure Gaussian noise.

[0066] Further, the conditional guidance generation mechanism uses the data of one or more dimensions in the multi-dimensional design scheme obtained in the design scheme process as a guidance condition to guide the output of the next stage model; the multi-dimensional design scheme includes a multi-dimensional preliminary design scheme and a multi-dimensional modified design scheme.

[0067] Data source and dataset production:

[0068] The dataset for model training and verification of the present application is from the multi-dimensional data of 1001 parks in 116 cities in China. Each park dataset contains real scene images, site function zoning, corresponding park road network, hand-drawn data, and park boundary data. The park real scene data is from World Imagery images. The site function zoning, corresponding road data, and park hand-drawn scheme set are drawn by 10 team members according to the corresponding park, and are digitized after scanning; the site function zoning is classified into hard, green, and water; the park road network data is marked with different road levels and directions with different widths and directions of black and white; the hand-drawn scheme data is drawn by the team members with colored pencils to draw the trees, lawns, roads, water areas, and existing buildings within the park. The park boundary data comes from the park fence line or land boundary.

[0069] The above data is uniformly processed and stored in the database. Renamed as image-xxxx and constructed a multi-dimensional database. The park real scene image data (RGB) and the hand-drawn scheme data (Draw) are in RGB png image format; the site function zoning (Class) is in classification tif format; the traffic flow line data (Grey) is in grayscale tif format data, and the mask data (Mask) is in binary png image format data.

[0070] MCDM model training method:

[0071] MCDM trains the special models of the three stages based on the Pycharm platform. The training is run on NVIDIARTX3060 GPU. After repeated iterative training and parameter adjustment comparison test, the models of the three stages are converged in about 300 rounds of models, and the final model with better effect is obtained through comparison and verification.

[0072] Application of MCDM in landscape planning:

[0073] Case overview:

[0074] The Lushang Ecological Park is located in a low hilly plain area, belonging to a cold temperate continental monsoon climate. The design area covers an area of about 8 hectares. An artificial lake is expected to be reclaimed in the site, but the area, shape, and location are uncertain. The existing internal ridge road and idle factory building in the site will not be retained.

[0075] The first stage "preliminary design" generates the effect:

[0076] In the preliminary design stage, the invention first collects the latest site real image (RGB) and traffic flow line data (Grey) of the case site by the unmanned aerial vehicle; then, according to the park real image, the design range data (Mask) and the functional partition data (Class) are made (such as Figure 5 ). The above site status data set is input into the first stage model to generate 20 different schemes, and the generation effect is as follows Figure 6 From the figure, it can be seen that each multi-dimensional preliminary design scheme generates a real scene scheme and corresponding road and functional partition according to the real park scene, and has consistency.

[0077] The second stage "scheme modification" generates the effect:

[0078] In the scheme modification stage, according to the opinions of Party A, No. 11 scheme in the preliminary design is selected as the preliminary design selected scheme, on the basis of which the proportion of hard area is manually reduced according to the requirements of Party A, and the position of blue and green space in the park is roughly delimited to form a new functional partition constraint image, finally the above constraint conditions are input into the second stage model to regenerate 20 sets of modified schemes (such as Figure 7 ).

[0079] The third stage "general plan design" generates the effect:

[0080] According to the selection of the second stage generated scheme by Party A, the park form and functional partition after the park scheme modification that best meet the needs are determined, combined with the model of the third stage "general plan design", 20 sets of modified refined hand-drawn plan scheme design drawings without road network constraint and 20 sets of modified refined hand-drawn plan scheme design drawings with road network constraint are finally generated (such as Figure 8 ). Party A finally selects No. 7 scheme under the condition of road network constraint, which is provided for the landscape architect to further refine the general plan design drawing.

[0081] In the application case, the first stage MCDM performs multiple conceptual division designs on the blue-green space in the design area, and generates corresponding multi-dimensional preliminary design schemes and road networks. From the results, it can be seen that the schemes have diversity and consistency between the output multi-dimensional data. However, the randomness of the schemes in this stage is high, and the most suitable scheme needs to be found from multiple different random schemes. In the second stage, the function division guided by manual division is added, so that the MCDM can modify the design on this basis. From the generated modified scheme, it can be seen that the design has obvious functional consistency. In the third stage, based on the multi-dimensional design data output by the second stage, the MCDM performs planar graph refinement design. It can be seen that the total plan design graph generated by the multi-dimensional modified design scheme and the function division guide has high quality and diversity. In addition, the additional road network guide can more finely meet the needs of Party A.

[0082] Comparison and evaluation of the present application and prior art:

[0083] In order to standardize the test of each generation stage as much as possible, evaluate the generation effect and advantages of the model, the present application combines the commonly used GAN algorithm generation model (Pix2Pix, CycleGAN, StyleGAN), Stable Diffusion, uses the same sample for training, and generates pictures for comparison, and uses multiple different indicators to comprehensively evaluate their performance in landscape garden plan design (Table 1).

[0084] Table 1 Comprehensive evaluation of the applicability of different picture generation models in landscape garden design

[0085]

[0086] (Note: "√" means that the function is available, "△" means that the function is limited, and "×" means that the function is not available. * The model can be trained with different samples and rounds to realize the function)

[0087] Figure 9 The out-of-picture examples of different picture generation models in landscape garden design are shown in the table. Through the comparison effect of different generation models, it can be seen that: 1) The preliminary design scheme generation effect of the traditional adversarial neural network Pix2Pix and CycleGAN model is poor, and the out-of-picture result of the same round model is unique. Although the StyleGAN model can distinguish and control different styles in the training set, the degree of diversification of the generated scheme is low, and it cannot meet the actual constraint requirements of design. 2) The preliminary design scheme generation effect of the SD model is good, and the generated result has diversity but lacks multi-dimensional data input and output. 3) The MCDM model has the diversity of output results and dimensions, and has good performance in urban park planning and design.

[0088] Example 2:

[0089] The garden plane design system based on a multi-dimensional constraint diffusion generation model comprises a multi-dimensional constraint diffusion generation model, an information input module and an information analysis output module.

[0090] The multi-dimensional constraint diffusion generation model realizes model training and scheme design processes of multi-dimensional data; the design scheme process comprises a preliminary design stage, a scheme modification stage and a total plane design stage; the multi-dimensional constraint diffusion generation model trains to obtain a first model in the preliminary design stage, trains to obtain a second model in the scheme modification stage and trains to obtain a third model in the total plane design stage.

[0091] The information input module can pre-process (such as adjusting the design range) and input the obtained garden design site data.

[0092] In the preliminary design stage, the information analysis module inputs the garden design site data into the first model to obtain a plurality of different multi-dimensional preliminary design schemes; in the scheme modification stage, a condition-guided generation mechanism is introduced to take the data of a certain dimension in the selected multi-dimensional preliminary design scheme as a constraint condition, modify the constraint condition as a guide condition as required, input the guide condition and the data of the remaining dimensions in the multi-dimensional preliminary design scheme into the second model to obtain a plurality of different multi-dimensional modified design schemes; in the total plane design stage, the condition-guided generation mechanism is used to take the data of a plurality of dimensions in the selected multi-dimensional modified design scheme generated in the scheme modification stage as optional guide conditions, input the guide conditions of the selected plurality of dimensions and the site real scene data in the selected multi-dimensional modified design scheme into the third model to obtain a plurality of different total plane design schemes, and finally select a total plane design scheme meeting the requirements. In the third model, the site real scene data only provides surrounding environment information.

[0093] The innovation of the present application lies in proposing and developing a multi-dimensional constraint diffusion generation model and system, and verifying that the model and system can help designers more purposefully solve design problems encountered in the "three stages" of landscape garden plane design, thereby reducing the work burden and improving the efficiency. The research results show that: 1) in the preliminary design stage, the designer can input the design site conditions obtained in the preliminary research stage into the model to quickly generate diversified multi-dimensional preliminary design schemes; in the scheme modification stage, the designer can customize the modified functional partition according to the modification opinions of Party A and quickly generate a plurality of plane real scenes, thereby greatly improving the efficiency of communication and feedback; in the total plane design stage, the last stage scheme, in addition to optional functional partition map and road network map data, can be used as guide data to generate a hand-drawn style total plane map. 2) The quality of the generated design scheme is high, compared with traditional non-professional picture generation models, and a series of human-computer collaborative design features such as multi-dimensional data input and output, scheme diversification generation and flexible control of generated content can be realized.

Claims

1. A landscape planar design method based on a multi-dimensional constraint diffusion generation model, characterized in that, The method comprises the following steps: (1) constructing a multi-dimensional constraint diffusion generation model to realize model training and design scheme process of multi-dimensional data; The design scheme process comprises a preliminary design stage, a scheme modification stage and a general plan design stage; The multi-dimensional constraint diffusion generation model trains a first model in the preliminary design stage, trains a second model in the scheme modification stage and trains a third model in the general plan design stage; (2) in the preliminary design stage, inputting garden design site data into the first model to obtain a plurality of different multi-dimensional preliminary design schemes; (3) in the scheme modification stage, introducing a conditional guided generation mechanism to take the data of a certain dimension in the selected multi-dimensional preliminary design scheme as a constraint condition, modifying the constraint condition as a guide condition as required, inputting the guide condition and the data of the remaining dimensions in the multi-dimensional preliminary design scheme into the second model to obtain a plurality of different multi-dimensional modified design schemes; (4) in the general plan design stage, using the conditional guided generation mechanism to take the data of a plurality of dimensions in the selected multi-dimensional modified design scheme generated in the scheme modification stage as optional guide conditions, inputting the guide conditions of the selected plurality of dimensions and the site real scene data in the selected multi-dimensional modified design scheme into the third model to obtain a plurality of different general plan design schemes, and finally selecting a general plan design scheme meeting the requirements.

2. The garden planar design method based on a multi-dimensional constraint diffusion generation model according to claim 1, characterized in that, The step (1) comprises: (1.1) in the model training process and the design scheme process, the input multi-dimensional data formula is expressed as: y0 = [RGB, Grey, Class, Draw] e R C×H×W wherein the input multi-dimensional data y0 is combined into a multi-channel conditional tensor R after tensorization, and the shape of R is channel number C x height H x width W; RGB represents a site real scene, Grey represents a site traffic flow line, Class represents a present situation functional partition, and Draw represents a site real scene hand-drawing; Noise predicted by the model at time step t The expression for the noise is wherein f θ is the trained model prediction network, is the element-wise multiplication, M is the mask tensor, x cond is the multi-dimensional conditional input, γ t is the noise scheduling coefficient, ∈ is the standard Gaussian noise, is the noisy tensor data; (1.2) in the model training process, the loss function L is calculated according to the following formula: wherein, is the expected value, represents the probability weighted average of the variable marked by its subscript, t is the time step, y0 is the input multi-dimensional data, ∈ k is the input real noise, k is each input multi-dimensional data, λ k is the proportion weight of each channel tensor, if the input is a guide condition, the corresponding proportion weight λ k = 0, is the predicted noise obtained by the model after t time steps of reasoning, ∈ k is the real noise of the initial input, and p is a positive integer.

3. The garden planar design method based on a multi-dimensional constraint diffusion generation model according to claim 1, characterized in that, The first model generates a random scheme in the preliminary design stage, the second model generates a scheme combined with artificial intervention in the scheme modification stage, and the third model generates a refined scheme in the general plan design stage; the refined scheme is a scheme capable of reflecting a specific style.

4. The garden planar design method based on a multi-dimensional constraint diffusion generation model according to claim 1, characterized in that, The multi-dimensional data used in the model training process comprises a site design range, a site real scene, a site traffic flow line, a present situation functional partition and a site real scene hand-drawing; in the design scheme process, the garden design site data comprises a site design range, a site real scene, a site traffic flow line and a present situation functional partition.

5. The garden planar design method based on a multi-dimensional constraint diffusion generation model according to claim 1, characterized in that, In the process of inputting multi-dimensional data into the multi-dimensional constraint diffusion generation model, a mask mechanism is used to demarcate a redrawing design area, a multi-dimensional constraint diffusion generation model is used to forward diffuse a noise adding mechanism, and after noise adding, the multi-dimensional constraint diffusion generation model is inputted again; the noise adding mechanism is a mechanism for gradually adding Gaussian noise to a target area with the passage of time step t until the state is indistinguishable from pure Gaussian noise.

6. The method of claim 1, wherein, The conditional guided generation mechanism refers to taking the data of one or more dimensions in a multi-dimensional design scheme obtained in the design scheme process as a guide condition to guide the output of a next stage model; The multi-dimensional design scheme comprises a multi-dimensional preliminary design scheme and a multi-dimensional modified design scheme.

7. A landscape plan design system based on a multi-dimensional constraint diffusion generation model, characterized by, The system comprises a multi-dimensional constraint diffusion generation model, an information input module and an information analysis output module; The multi-dimensional constraint diffusion generation model realizes model training and scheme design process of multi-dimensional data; The design scheme process comprises a preliminary design stage, a scheme modification stage and a general plane design stage; The multi-dimensional constraint diffusion generation model trains to obtain a first model in the preliminary design stage, trains to obtain a second model in the scheme modification stage and trains to obtain a third model in the general plane design stage; The information input module can preprocess and input obtained garden design site data; In the preliminary design stage, the information analysis module inputs the garden design site data into the first model to obtain a plurality of different multi-dimensional preliminary design schemes; In the scheme modification stage, a conditional guided generation mechanism is introduced to take data of a certain dimension in the selected multi-dimensional preliminary design scheme as a constraint condition, modify the constraint condition as a guide condition on demand, input the guide condition and data of the remaining dimensions in the multi-dimensional preliminary design scheme into the second model to obtain a plurality of different multi-dimensional modified design schemes; In the general plane design stage, the conditional guided generation mechanism takes data of a plurality of dimensions in the selected multi-dimensional modified design scheme generated in the scheme modification stage as optional guide conditions, inputs the guide conditions of the selected plurality of dimensions and the site real scene data in the selected multi-dimensional modified design scheme into the third model to obtain a plurality of different general plane design schemes, and finally selects a general plane design scheme meeting the requirements. In the model training process and the design scheme process, the input multi-dimensional data formula is expressed as:

8. The garden plan design system based on a multi-dimensional constraint diffusion generation model according to claim 7, characterized in that, In the model training process and the design scheme process, the input multi-dimensional data formula is expressed as: y0 = [RGB, Grey, Class, Draw] e R C×H×W In the process of inputting the multi-dimensional data into the multi-dimensional constraint diffusion generation model, a mask mechanism is used to demarcate a redrawing design area, and a multi-dimensional constraint diffusion generation model is used to diffuse a noise adding mechanism forward, and the multi-dimensional constraint diffusion generation model is inputted after noise adding processing; the noise adding mechanism is a mechanism for gradually adding Gaussian noise to the target area with the passage of time step t until the state is not different from pure Gaussian noise. Noise predicted by the model at time step t The expression for the noise is wherein f θ is a trained model prediction network, is an element-wise multiplication, M is a mask tensor, x cond is a multi-dimensional conditional input, γ t is a noise scheduling coefficient, ∈ is a standard Gaussian noise, is a noisy tensor data; In the model training process, a loss function is designed The calculation formula is as follows: wherein, is the expected value, represents the probability weighted average of the variable marked by its subscript, t is the time step, y0 is the input multi-dimensional data, ∈ k is the input real noise, k is each input multi-dimensional data, λ k is the proportion weight of each channel tensor, if the input is a guide condition, the corresponding proportion weight λ k = 0, is the predicted noise obtained by the model after t time steps of reasoning, ∈ k is the real noise of the initial input, and p is a positive integer.

9. The garden plan design system based on a multi-dimensional constraint diffusion generation model according to claim 7, wherein, The conditional guided generation mechanism refers to taking data of one or more dimensions in the multi-dimensional design scheme obtained in the design scheme process as a guide condition to guide the output of the next stage model; 10. The garden plan design system based on a multi-dimensional constraint diffusion generation model according to claim 7, characterized in that, The multi-dimensional design scheme comprises a multi-dimensional preliminary design scheme and a multi-dimensional modified design scheme. ​

Images