How to generate a museum floor plan: LoRA for diffusion models, or GANs?

Abstract

This study utilized a comparative methodology to explore the generative design of museum floor plans, addressing the complexity of meeting curatorial, designer, and visitor demands. A dataset of 263 museum plans was curated, and four generative methods―LoRA for diffusion models, Pix2Pix, CycleGAN, and the generative segmentation model―were trained. The results were evaluated by FID, SSIM, and MAE metrics for image fidelity and diversity. The pixel-based space syntax method was applied to evaluate the connectivity of the functional space, and expert scoring was combined to assess whether the generated plans met the layout characteristics of museums. Findings are: First, Pix2Pix led in terms of the image structural similarity metric SSIM and diversity metric FID but scored lowest in the image detail reconstruction metric MAE. Second, space syntax analysis revealed that the generated layouts had greater exhibition space connectivity but lower public space connectivity than the original images. Third, the expert grading results indicate that the outputs of LoRA better align with the professional expectations of designers regarding functional composition and spatial organization. These findings highlight the potential of generative models in museum design, particularly when enhanced by space syntax evaluations to improve spatial intelligibility and meet diverse stakeholder needs.

Graphical abstract

Keywords

Museum floor plan design / Intelligent generative design / Generative adversarial network (GAN) / Diffusion models / Space syntax

1 Introduction

Museums are critical landmarks and knowledge repositories within city environments (van Aalst and Boogaarts, 2002). A museum's floor plan represents the configuration of different functional spaces and exhibition elements. The connection or separation between exhibition spaces, the ordering of elements, and the accessibility patterns of the space form visitors' perceptions and shape their understanding (Wineman and Peponis, 2010). Thus, museum floor plan design significantly affects how visitors navigate and enjoy exhibits, and research on floor plan design methods could help enhance museum cultural communication (Min et al., 2023). The design of a museum is a complex process that need to balance multiple factors: the optimization of exhibition arrangements, the assurance of a good visiting experience, and the rational configuration of functional spaces. Floor plan design in museums necessitates a profound understanding of topological spatial configurations, yet the process remains repetitive and laborious for designers (Karadag et al., 2022; Ko et al., 2023). The traditional floor plan design method in architecture often involves sketching on tracing paper to explore different design options and address programmatic relationships in 2D and 3D spaces (Stitt, 1984). However, this method is typically iterative and time-consuming, requires significant problem-solving, and might overshadow the design's creative aspects (Tang et al., 2024). To address these concerns, generating multiple design alternatives and utilizing logical or abductive reasoning has proven effective in navigating the complexities of floor plan design. Consequently, automation in floor plan design with computer assistance has emerged as a solution to these nonlinear design challenges (Rahbar et al., 2022).

Recently, the new computational revolution, artificial intelligence (AI), has significantly contributed to the development of automated floor plan design for automating mundane tasks, improving design quality and efficiency, and reducing design errors (Ko et al., 2023). Among those AI-based floor plan design methods, Pix2Pix, CycleGAN, and the diffusion model have received extensive attention in previous research. For instance, Pix2Pix could be applied to generate energy-efficient residential building floor plans with the Solar Decathlon (SD) entry sample set (Wan et al., 2022). Moreover, combining Pix2Pix and agent-based models could increase the computational accuracy in generating synthetic architectural layouts (Rahbar et al., 2022). In addition, several studies have compared the generation results of the Pix2Pix and CycleGAN models to explore the applicability of these two models in different generation scenarios (Chen et al., 2023; Zhao et al., 2021) or with different input conditions (Güzelci et al., 2024). Furthermore, Zeng et al. (2024) reported that for nonprofessional users, generation models based on GANs and GNNs presented limited flexibility in terms of input conditions and low controllability in outputs. Specifically, nonprofessional users lack the expertise required to interpret the abstract design inputs typically used by professionals. However, they still need some customization options, such as adjustable room sizes, layouts, and footprint of the overall residence, to tailor the design to their own needs; thus, they proposed a diffusion model-based generation approach to implement multicondition model input and enhance the model's input flexibility. Moreover, Yang et al. (2022) trained a segmentation model using semantics and topology-based methods, enabling the transformation of input residential raster floor plans into vectorized floor plans and ultimately generating a structured architectural interior space. Owing to the simplicity and availability of well-structured residential datasets driven by the commercial nature of housing markets (Merrell et al., 2010; Wan et al., 2022), these previous applications of floor plan generative design focused mainly on residential buildings and achieved impressive progress (Keshavarzi and Rahmani-Asl, 2021; Liu et al., 2024; Wang and Duan, 2023; Weber et al., 2022).

In contrast, the museum's function, form, and spatial configuration significantly differ from those of ordinary residences. Museums should not only meet the basic functions of exhibition and collection but also create a comfortable and immersive experience for visitors. Thus, the museum's floor plan is intrinsically complex and includes specialized functions such as exhibition space, public space, storage space, and office space (Hecht et al., 2020; Rani et al., 2023). Owing to the lack of comprehensive datasets of museum floor plans, only a few studies have explored AI-based museum floor plan generation, and existing studies have discussed mainly the generation of furniture arrangements such as exhibition cabinets in the exhibition space of museum floor plans (Min et al., 2023; Tang et al., 2024), but have not focused on the architectural floor plan layout of multiple exhibition halls and public spaces of the museum. Moreover, these studies adopted only the GAN model to generate floor plans and did not compare the quality of the results generated by different AI models.

Thus, this research adopted four commonly used automated floor plan design methods (e.g., LoRA for the diffusion model, Pix2Pix, CycleGAN, and the generative segmentation model) to generate museum floor plans. The generated outcomes were analyzed and compared from the perspective of museums as a distinct building type, and the strengths and weaknesses of each method were identified in addressing the specific demands of museum layouts. Simultaneously, this research focuses on the spatial configuration of different exhibition halls, atriums, and corridors in architectural dimensions but not on the furniture layout inside the exhibition hall. The primary objectives of this research are as follows:

(1) To collect data and construct a comprehensive dataset of museum floor plans representing the unique spatial and functional characteristics of this building type.

(2) To employ several state-of-the-art methods (e.g., LoRA for the diffusion model, Pix2Pix, CycleGAN, and the generative segmentation model), which have been proven effective in floor plan generation, to train and generate results based on the curated museum floor plan dataset, yielding outcomes of acceptable quality.

(3) To innovatively combine the analysis of spatial characteristics based on space syntax and image quality assessment (IQA) metrics to evaluate and compare the performance of the generative model in museum floor plan design from the perspective of museums as a distinct building type.

2 Related work

2.1 Museum floor plan design

A floor plan is a fundamental tool for architects when designing buildings. In traditional architectural design, it typically takes a team of architects several days or even months to optimize and complete a building's floor plan. However, generative design research for floor plans aims to significantly shorten the time required to create architectural layouts. In this process, the role of the architect shifts from handling every minute detail to focusing on controlling key nodes of the floor plan. This transition can greatly increase the efficiency of architects. Generative floor plan design has already undergone significant research developments (Wang and Zhang, 2020), such as the shape grammar-based floor plan generation method proposed by Stiny (Eilouti, 2019; Stiny, 1980), which operates on the basis of given boundaries. Kahng (2000) introduced a spatial design method based on components. Museums are typically planned and constructed by the government and are open to the public for exhibitions and activities. The functionality of a museum makes it a complex building constructed collaboratively by multiple stakeholders. This complexity may explain why research on floor plan generation for museums is limited. Previous studies have applied generative convolutional generative adversarial networks (GCGANs) to the layout of exhibits in museum exhibition halls (Min et al., 2023; Tang et al., 2024). However, these studies focus solely on the arrangement of exhibits in exhibition spaces and do not address the complex relationships between different functions within the museum. Consequently, research on the automated generation of entire museum floor plans is lacking.

2.2 Generative models adopted in floor plan generation

There is a significant amount of research on floor plan generation, with most studies focusing on two primary categories: diffusion models and generative adversarial network (GAN) models. Therefore, this study primarily explores these two categories. Specifically, four methods are selected for analysis, LoRA based on diffusion models, Pix2Pix and CycleGAN models from the GAN category, and generative segmentation models, which replace the output of the GAN models with channel outputs. These methods serve as the basis for the analysis and research presented in this paper.

2.2.1 Diffusion models adopted in floor plan generation

Diffusion models have become the leading algorithm for generating high-quality floor plan images and generate images by adding noise and denoising. Compared with traditional generative algorithms such as GANs, diffusion models yield superior results in terms of image fidelity and detail preservation, particularly in complex architectural floor plans (Ho et al., 2020). However, training and fine-tuning diffusion models are computationally expensive and require large-scale datasets and significant GPU resources. In previous studies, the diffusion model has been used mainly in the exploration of learning (Su et al., 2024) and generating (Zeng et al., 2024) for residential floor plan design because of the massive database of residential floor plans. In contrast, it is difficult for museums to train directly through diffusion models because of their limited number of cases.

To mitigate computational costs, parameter-efficient methods for fine tuning such as low-rank adaptation (LoRA) have been adopted. LoRA enables the fine-tuning of large diffusion models (e.g., stable diffusion) for specialized tasks with limited data, making them accessible to smaller research teams (Rombach et al., 2022). In recent years, LoRA has been adopted in shear wall structure layout generation (Zhou et al., 2024), urban design image generative design (Cui et al., 2024), local traditional building facade generative design (Xu et al., 2024), and standard nursing unit floor plan generative design (Han and Chen, 2024). However, few studies have explored museum floor plan generative design based on LoRA for diffusion models.

2.2.2 Generative adversarial networks (GANs) for floor plan generation

The generative adversarial networks (GANs) proposed by Goodfellow et al. have revolutionized image generation by establishing a competitive framework between two neural networks: a generator that synthesizes data and a discriminator that evaluates authenticity. This adversarial training paradigm enables GANs to produce high-fidelity outputs that preserve essential features of input data variations, making them particularly effective for image synthesis tasks (Goodfellow et al., 2020). Building on the GAN framework, Pix2Pix pioneered conditional image-to-image translation by training on paired datasets. By incorporating pixel-level supervision, the model learns to map specific input conditions (e.g., edge maps) to realistic output images (Isola et al., 2017). Moreover, to address scenarios lacking paired data, CycleGAN introduces cycle-consistent adversarial training. By enforcing bidirectional mappings between domains, the model learns intrinsic feature relationships without requiring exact input–output pairs (Zhu et al., 2017). GAN-based Pix2Pix and CycleGAN methods have been adopted widely in architectural floor plan generation because of their ease of training and ability to produce meaningful results, even with small-scale datasets. Specifically, Li and Xu (2022) employed Cycle-GAN to translate sketches into building images using a dataset of 200 samples, whereas Kösenciğ et al. (2024) applied Pix2Pix to generate residential floor plans with structural elements from a set of 100 samples. Moreover, using a dataset of 120 samples, Zhao et al. (2021) compared the applicability of Pix2Pix and CycleGAN in generating an emergency room floor plan.

The results of floor plan generation, being a type of image, can naturally be classified as an RGB pixel image, similar to the input and output of models such as Pix2Pix and CycleGAN (Isola et al., 2017; Zhu et al., 2017).However, floor plans, particularly spatial layout plans that describe the spatial distribution of specific functional areas, primarily represent the spatial positioning of various functions. This spatial representation closely resembles segmentation. Therefore, while maintaining the fundamental generator-discriminator architecture, replacing the input and output with a channel-based representation can lead to different effects. Previous studies have explored the use of segmentation models as a generative approach. Yang et al. (2022) focused on residential interior spaces and trained a segmentation model to convert input raster floor plan images into vectorized floor plans, thereby generating a three-dimensional architectural space based on the input raster floor plan. Furthermore, Pizarro et al. (2023) extended this model from individual residential floor plans to unit-based residential floor plans, making it adaptable to a broader range of scenarios. While these studies further support the effectiveness of channel-based feature learning, as exemplified by segmentation, few studies have compared the generative results of channel-based segmentation algorithms and RGB pixel image-based GAN algorithms in terms of floor plan generation.

2.3 Evaluation metrics for the museum floor plan generation results

Choosing appropriate evaluation metrics is important for comprehensively assessing the performance of floor plan generative models. Both quantitative and qualitative metrics were used to evaluate the quality of the generated images and their underlying spatial configurations.

2.3.1 Quantitative evaluation of generative floor plan image quality

Visual quality assessment of generated floor plans employs established computer vision metrics to provide objective performance measures. For instance, the Frechet inception distance (FID) measures the statistical similarity between generated and reference images through learned feature distributions (Heusel et al., 2017); the structural similarity index measure (SSIM) evaluates perceptual quality by analyzing luminance, contrast, and structural relationships within the images (Wang et al., 2004); and the mean absolute error (MAE) quantifies the average absolute difference between generated outputs and target images at the pixel level (Zhang et al., 2018). For generative architectural images, the FID metric has been adopted to evaluate the quality of generative architectural shape sketches (Qian et al., 2022). The SSIM metric can be adopted to evaluate the structural similarity of generative classroom layouts (Karadag et al., 2022), predict historical architectural facades (Güzelci et al., 2024), and predict the daylight performance of general floor plans (He et al., 2021). In addition, the combined application of the FID, SSIM, and MAE metrics enables a comprehensive evaluation of the quality of generated architectural images (Nourkojouri et al., 2023; Zeng et al., 2024).

2.3.2 Qualitative evaluation via expert grading

While quantitative image quality evaluation metrics provide objective measures, expert grading remains indispensable for evaluating the perceptual realism, functional adequacy, and human acceptance of AI-generated designs with specific professional domain knowledge (Fu et al., 2023). To mitigate the subjectivity in expert scoring, a few studies have used methods such as the analytic hierarchy process (AHP) to construct expert evaluation frameworks (Doğan et al., 2022; Kim, 2021). In addition, Fatemeh Mostafavi et al. explored hybrid evaluation methods in which expert judgments are supplemented with quantitative metrics and reported that while AI-generated plans scored highly in terms of computational efficiency metrics, expert reviewers often penalized them for lacking contextual sensitivity (Mostafavi et al., 2025).

2.3.3 Combined quantitative and qualitative evaluation via space syntax

The spatial layout of a museum and the correlation of different functional spaces play crucial roles in realizing curatorial intent, creating a good visitor experience, and further enhancing the ability of a visitor to interpret a museum's historical and cultural information (Wineman and Peponis, 2010). Thus, the assessment should not be limited to the quality of the generated images, and the museum's spatial configuration and the results' objectivity should also be maintained. The spatial analysis method space syntax proposed by Bill Hillier is a theoretical and analytical framework used to study the spatial configurations of architectural environments. With respect to museum floor plan design, space syntax could provide a rigorous methodology for evaluating how the arrangement of spaces influences visitor movement and overall experience (Hillier and Tzortzi, 2006; Lu, 2010).

Space syntax operates as a hybrid analytical framework that integrates quantitative metrics with qualitative interpretation. At its core, it employs mathematical graph theory (Turner et al., 2001) and computational algorithms to generate objective spatial indices―such as connectivity, integration, and intelligibility values (Hillier and Hanson, 1984) ―which quantitatively describe topological relationships in built environments. However, as Hillier emphasizes, these metrics only attain meaning when contextualized through socio-spatial theory, requiring a qualitative examination of how numerical results correlate with human behavior patterns or design intentionality (Hillier, 1999; Wineman and Peponis, 2010). Karimi (2018) thus characterized space syntax as a quantitative toolkit for qualitative questions, bridging positivist measurement with a phenomenological understanding of spatial culture.

Furthermore, most existing spatial syntax studies have used Depthmap software to analyze the vectorized graphic of building planes in dxf format (Hegazi et al., 2022; Wong, 2014), and few studies have focused on the spatial syntax analysis of planes based on pixel images; however, pixel images are the basic format for generating planes for AI models. Unlike vector graphics represented by geometric shapes that can be conveniently mathematically expressed, pixel-based images are graphical outcomes composed of pixel arrays. Consequently, spatial syntax computation in pixel-based environments becomes computationally intensive. The calculation of visual occlusion in spatial syntax is similar to that of collision detection algorithms in computer graphics, where the ray casting method―a common geometric computation technique in collision detection―can be employed to alleviate computational demands in this context (Appel, 1968; Wang et al., 2012).

In summary, while diffusion models based on LoRA and GANs have advanced floor plan generation, their application to museums remains limited. Museums' complex spatial and functional requirements pose unique challenges; however, existing research focuses only on furniture arrangements within exhibition spaces and neglects entire museum layouts. Additionally, comparative studies evaluating generative models such as LoRA, Pix2Pix, CycleGAN, and segmentation models for museum design are lacking. Therefore, this research addresses these problems by developing a museum floor plan dataset, employing state-of-the-art generative methods, and combining space syntax with image quality analysis to assess model performance in museum layout design to summarize the capability of different generative models in museum floor plan generation.

3 Methodology

3.1 Workflow and framework

This study employs four methods, namely, stable diffusion with LoRA, Pix2Pix, CycleGAN, and the generative segmentation model, within a two-stage model to generate floor plan layouts for public spaces in museums, with a focus on areas accessible to visitors. The aim of this research is to provide faster, more efficient, and more diverse design possibilities for museum floor plan layouts. The framework is divided into three main parts: data preprocessing, training, and evaluation. It integrates four methods and two-stage model training while analyzing the trained models and evaluating the generated results using space syntax analysis. Figure 1 illustrates the research framework.

Moreover, the comparative methods employed in this study, LoRA, Pix2Pix, CycleGAN, and the generative segmentation model, must meet the following criteria. First, the models should be trained effectively on the dataset we created. Second, the generative process must allow designers to input predefined boundaries, as boundaryless generation offers limited utility for practical design applications. Finally, the generated results must feature clear functional boundaries to enable a comprehensive comparison of the floorplans produced by different methods.

In the data preprocessing stage, floor plan data from 263 museums were collected and annotated according to the correspondence between spatial functions and colors, enabling clear functional distinctions within the museum layouts. To ensure that image features were accurately learned during training, data augmentation techniques were applied.

During the training phase, image features were extracted from the input dataset, and latent features were learned and recorded through the model training process. The trained models leveraged these learned features to generate output images on the basis of various input image controls. The two-stage model training framework divides the generation process into two sequential tasks: first, the layout of functional areas is generated, and second, the walls that represent the spatial connections between these functional areas are generated.

In this evaluation framework, the assessment is conducted from two perspectives: the layout generation itself and space syntax, which represents spatial topological relationships. These two aspects are further divided into qualitative and quantitative analysis dimensions. The quantitative analysis of layout generation employs conventional image generation metrics, including the FID, SSIM, and MAE. For space syntax quantitative evaluation, spatial connectivity within individual spaces serves as the primary measurement. The qualitative analysis of layout generation involves subjective evaluations from 21 invited designers or engineers, with statistical aggregation of their assessments. The qualitative analysis of space syntax relies primarily on the author's interpretation of space syntax heatmaps for analytical evaluation.

3.2 Data preprocessing

3.2.1 Data capture and augmentation

While the specific exhibits vary across museums, the fundamental logic of displaying artifacts and facilitating visitor engagement remains consistent across different institutions. This study focuses on generating floor plan layouts for public spaces in museums to explicitly address functional areas such as corridors, foyers, permanent exhibition halls, temporary exhibition halls, and lecture halls―spaces accessible to visitors. Administrative and logistical spaces, including offices for museum staff and storage areas for artifacts, differ significantly in their requirements across museums and lack standardized functional abstraction. Moreover, these spaces are typically closed to the public and hold little relevance to the study's focus on visitor experience. Therefore, such areas are uniformly labeled as other functions in the dataset. Given the variation in museum size, imposing strict scale requirements on the dataset would limit its size and hinder its utility for training. Since the study does not emphasize scale-specific analysis, the dimensional aspect of museums is not extensively discussed.

According to these considerations, we extensively investigated museum floor plan data and ultimately selected 263 museum floor plan datasets on the basis of the following selection criteria. This dataset has been made open-source on GitHub and primarily includes the original floor plan data and the color-labeled training data prepared for this study and generated results (Song, 2025).

(1) Floor plans must clearly label room names to represent their functions.

(2) The walls and doors must be distinctly represented to illustrate spatial connections.

(3) Boundary information must be precise and clean.

(4) To ensure realistic representations, functional room boundaries should include diverse shapes.

To enhance the model's generalization ability and prevent overfitting, the dataset underwent augmentation processes such as scaling, rotation, and mirroring. These augmentations significantly expanded the dataset, providing a broader variety of architectural boundaries to improve the adaptability of the model to diverse floor plans.

3.2.2 Two-stage processing

Typically, model training follows an end-to-end approach, which means that the process involves no intermediate steps or additional processing. The input is fed directly into the model, and the result is generated after a single pass through the network without requiring further adjustments or refinements. This streamlined approach allows the model to learn from raw data and produce an output in one step, simplifying the overall workflow. However, for architectural floor plans, walls and functional rooms have vastly different semantic meanings. Therefore, the model is trained in two stages and requires corresponding dataset preprocessing. In the first stage, wall structures are excluded, and floor plans consist solely of colored blocks representing different functions (Güzelci et al., 2024). In the second stage, these functional blocks are combined with black pixels representing walls. The color coding for functional blocks includes distinctions such as exhibition halls, temporary exhibition halls, lecture halls, corridors, atriums, foyers, and other functions, ensuring clear semantic representation. Representative images from the dataset are shown in Fig. 2, which illustrates the composition and characteristics of the data used in this study. It includes examples from the raw data, the images used in the training set for Phase 1, and those employed in the training set for Phase 2. These examples demonstrate the progression and variations in data preprocessing and selection across different training phases, providing a clear visual representation of the dataset structure.

3.3 Training

3.3.1 Two-stage model

Initial attempts to train GAN-based models using an end-to-end method resulted in suboptimal performance. While the external boundaries of buildings were well recognized, the internal walls, which represent critical spatial connections, were poorly captured. This challenge arises from the semantic dissimilarity between black pixels (walls) and colored pixels (functional areas). Therefore, training was divided into two stages: the first stage focuses on learning functional layouts, and the second stage emphasizes the generation of wall structures and their spatial connections.

3.3.2 LoRA training

Training diffusion models with billions of parameters requires substantial computational resources, time, and extensive datasets. Given the modest size of the museum dataset (263 samples; detailed in Section 3.2.1), direct training of large models is impractical. LoRA provides an efficient fine-tuning method by leveraging pretrained large models and significantly reducing the number of trainable parameters. LoRA introduces a low-rank approximation to the parameter matrix, reducing computational complexity while preserving the model's structural integrity. The training loss of LoRA can be mathematically represented as follows:

\begin{array}{}\text{(1)}& W=W_0+\Delta W=W_0+AB=W_0+(d\times r)\times (r\times k)\mathrm{.}\end{array}

In this equation, W0 represents the original pretrained model weights. The parameter matrix of the original model has dimensions of d × k. LoRA introduces a parameter r to reduce the dimensionality of the trainable model matrix. The value of r typically ranges from 1 to 64, with 32 selected for training purposes in this study. The complete training architecture of the LoRA model that provides a detailed visualization of the structural design of the model is shown in Fig. 3. Table 1 outlines all the hyperparameters employed in the LoRA model and provides a detailed reference for understanding its configuration and optimization (Hu et al., 2021).

3.3.3 GAN training

3.3.3.1 Pix2Pix training

The Pix2Pix GAN consists of a generator and a discriminator. The generator creates new images while the discriminator evaluates their authenticity. The discriminator's loss is as follows:

\begin{array}{}\text{(2)}& L_D=-{\mathbb{E}}_{x\mathrm{,}y}[\log D(x\mathrm{,}y)]-{\mathbb{E}}_x[\log (1-D(x\mathrm{,}G(x)))]\mathrm{.}\end{array}

The role of the discriminator is to correctly identify real images as genuine and generate images as fake, which means that D(x, y) should be maximized, while D(x, G(x)) should be minimized. The generator's loss comprises two parts: L_GAN and L₁.

\begin{array}{}\text{(3)}& L_G=L_{GAN}+\lambda L_1\mathrm{.}\end{array}

The L_GAN component is relatively straightforward. It deceives the discriminator into believing that the generated images are accurate, thus maximizing D(x, G(x)).

\begin{array}{}\text{(4)}& L_{GAN}=-{\mathbb{E}}_x[\log D(x\mathrm{,}G(x))]\mathrm{.}\end{array}

The second part, L₁ loss, quantifies the pixel-wise difference between the real and generated images. The coefficient λ controls the proportion of this loss in the generator's total loss. In this study, consistent with the original Pix2Pix paper, λ is set to 100. The entire training framework of Pix2Pix is depicted in Fig. 4, highlighting the key structural components and their interconnections. All the hyperparameters utilized in the Pix2Pix model are systematically presented in Table 1, which provides a clear and structured summary of the parameters (Isola et al., 2017).

\begin{array}{}\text{(5)}& L_1={\mathbb{E}}_{x\mathrm{,}y}[\Vert y-G(x){\Vert }_1]\mathrm{.}\end{array}

3.3.3.2 CycleGAN training

CycleGAN and Pix2Pix are both types of adversarial networks, each comprising two main components: a discriminator and a generator. The construction of their loss functions is similar, whereas CycleGAN introduces an additional component known as the cycle consistency loss.

\begin{array}{}\text{(6)}& L_{CycleGAN}=L_{GAN}+\lambda L_{cycle}\mathrm{.}\end{array}

This additional loss, termed cycle consistency loss, serves a crucial purpose. It involves passing an input image through the generator to obtain a transformed image, which is then passed back through the generator to reconstruct the original input image. The process can be described by the following cycle consistency loss formula:

\begin{array}{}\text{(7)}& L_{cycle}={\mathbb{E}}_{x\mathrm{,}y}\left[\parallel x-G_{YX}(G_{XY}(x)){\parallel }_1+\parallel y-G_{XY}(G_{YX}(y)){\parallel }_1\right]\mathrm{.}\end{array}

The trained model, therefore, can cyclically transform images between the input and target domains, ensuring that the transformations are reversible and that the integrity of the original images is maintained. A comprehensive visualization of the CycleGAN training architecture is shown in Fig. 5, emphasizing its design and operational processes. A complete list of the hyperparameters used in CycleGAN can be found in Table 1, serving as a concise reference for the model's implementation (Zhu et al., 2017).

3.3.3.3 Generative segmentation model training

The segmentation-based approach represents an optimization and variant of the generative adversarial network (GAN) framework, which is tailored explicitly for generating museum floor plan results akin to semantic segmentation tasks. The overarching GAN architecture remains intact. However, unlike traditional image-generating GAN models, this method directly produces corresponding labels. That is, the output is not an RGB pixel image with three channels but rather a channel image with multiple channels representing different functional categories. The floor plan is subsequently generated on the basis of these labels, with each functional category mapped to its respective color. In semantic segmentation tasks, cross-entropy loss is a method of calculating loss on the basis of the difference between the predicted label probabilities and the true labels.

\begin{array}{}\text{(8)}& L_{\text{cross}-\text{entropy}}=-{\displaystyle \sum _{c=1}^C{y}_c\log ({\widehat{y}}_c)\mathrm{.}}\end{array}

In the above formula, y_c represents the indicator variable for category c in the true label. It is a one-hot vector that denotes the actual category. If the true category of the sample is c, then y_c = 1, otherwise y_c = 0. {\widehat{y}}_c refers to the predicted probability for category c in the model's probability distribution. This is the model's output probability for category c, typically obtained using the softmax function.

This method, which focuses on label prediction rather than direct image generation, employs a statistical loss approach that is highly suitable for the segmentation-based task of generating label predictions. The training architecture of the generative segmentation model is thoroughly presented in Fig. 6. Table 1 comprehensively details the hyperparameters applied in the generative segmentation model, highlighting their role in defining the model's structure and functionality (Pizarro et al., 2023).

All the experiments discussed in this section were conducted on an AutoDL cloud computing platform, leveraging a single NVIDIA 4090D GPU for both training and computation.

3.4 Evaluation and analysis

3.4.1 Generative image quality evaluation metrics

To validate the performance of the floor plan generative models in this research, three key metrics were considered for comparing the generated results with the ground truth images: the Frechet Inception Distance (FID), Structural Similarity Index Measure (SSIM), and Mean Absolute Error (MAE). The MAE quantifies the pixel-level accuracy of the generated images by measuring the average absolute difference between the generated and ground-truth images to reflect the model's ability to reconstruct fine details. The SSIM evaluates the perceptual quality by considering luminance, contrast, and structural similarity to ensure that the generated images maintain a high level of visual coherence with the ground truth (Güzelci et al., 2024; Karadag et al., 2022; Wang et al., 2004). The FID assesses the diversity and realism of the generated images by comparing the feature distributions of the generated and real images in the latent space of the InceptionV3 model, providing insight into the model's ability to produce diverse and realistic outputs. Together, these metrics offer a multidimensional analysis of the generative model's image quality performance.

3.4.2 Space syntax analysis based on pixels

This study employs visibility graph analysis from space syntax to evaluate global visibility and accessibility in architectural spaces. The analysis requires spatial positions to be defined on a grid to examine interspatial relationships (Turner et al., 2001). AI-generated pixel-based images inherently possess pixel grid structures, establishing a foundation for the visibility graph analysis of raster images. Spatial evaluation via a visibility graph relies on visibility and accessibility between grid positions, which are determined by obstructions such as walls or windows. Statistical assessment of each position is conducted by quantifying visible and accessible points within the spatial configuration, enabling comprehensive spatial evaluation. In this research, visibility and accessibility between pixels can be interpreted through black pixels representing walls in generated images. Thus, visibility graph evaluation can be effectively applied to the pixel-based architectural floor plan images produced in this study.

Unlike existing research focusing on spatial syntax analysis of vector graphics in formats such as DXF, this study proposes a spatial syntax analysis method based on pixel images to efficiently and conveniently analyze AI-generated pixel-based images.

Visibility analysis is a fundamental component of space syntax, which can be understood as the spatial perception of each point/pixel in each space and its direct visibility of surrounding areas. In such analyses, spatial barriers, specifically walls, play a critical role in influencing the final quantification of the visibility of each point/pixel. In traditional space syntax analysis, walls are represented by two endpoints, and the computational complexity, denoted O, is expressed as follows:

\begin{array}{}\text{(9)}& O=n\times m\times b\mathrm{,}\end{array}

where n × m represents the number of spatial grid units and b refers to the number of wall barriers. For the AI-generated pixel results, all the walls are represented as a collection of black pixels. The computational complexity of visibility analysis in this case is expressed as follows:

\begin{array}{}\text{(10)}& O=n\times m\times B\mathrm{,}\end{array}

where B denotes the number of black pixels representing walls, which is typically much larger than the original number of barriers b. In this case, the computational time becomes excessively long, making the analysis nearly impractical. Therefore, this study draws on the logic of the conventional ray-tracing algorithm (Kajiya, 1986). In this modified approach, each pixel emits B rays outward, and replaces the original logic of computing visibility between all pairs of pixels. The value of B represents the precision, with each ray being traced until it encounters a black pixel. All pixels traversed by the ray before it reaches a black pixel are considered unobstructed. In this study, B is set to 36, meaning that a ray is emitted every 10°. This method significantly reduces computational complexity and allows for faster generation of results. The algorithm mentioned above, which replaces obstacle detection in the original method with a ray-based approach, is shown in Fig. 7. This modification enhances the computational framework by introducing a more efficient mechanism for spatial analysis.

3.4.3 Expert grading

To enhance our evaluation framework, we incorporated an expert questionnaire survey for plan analysis involving 21 designers and architects. The questionnaire comprises four primary components, as shown in Fig. 8. The first question evaluates the rationality of functional distribution in generated plans, which is primarily represented by spatial allocations of the exhibition hall, the temporary exhibition hall, the lecture hall, and other functions. The second question assesses functional connectivity relationships, where are characterized by the configurations of the corridor, atrium, and foyer. The third question examines the clarity of functional boundaries, as defined by wall placements. The final question provides an overall assessment of plan generation quality. Each evaluation method includes four variant plans generated under identical boundary conditions (Fig. 9), enabling comparative scoring of different models across these criteria.

4 Results

The analysis of the results is divided into two sections: qualitative and quantitative findings. To ensure comprehensiveness, each segment encompasses evaluations of both the layout generation outcomes and space syntax-based spatial topological relationships. For quantitative analysis of layout generation, the FID, SSIM, and MAE metrics are primarily employed for scoring the methods. The space syntax component utilizes mainly connectivity scores between functional spaces for comparative assessment. The qualitative analysis of layout generation aggregates the evaluations from 21 designers and engineers regarding the performance of various methods. The qualitative analysis of space syntax is derived principally from the author's interpretation of the space syntax heatmap results.

4.1 Quantitative results

From the perspective of image structural information (SSIM), the GAN and LoRA methods achieved high scores, indicating that the spatial layout features of the generated images closely resembled those of the ground truth. These results are consistent with the effectiveness of the Pix2Pix model in generating structural features of architectural floor plans, as validated by previous studies (Karadag et al., 2022; Mostafavi et al., 2022), and further indicate that CycleGAN, the segmentation model, and LoRA suggest effective structure feature learning in museum spatial layouts. With respect to image detail reconstruction (MAE), Pix2Pix scored the lowest, whereas CycleGAN and the generative segmentation model achieved better scores. Cyclic consistency loss in CycleGAN facilitated broader learning across the dataset, whereas the channel output in the generative segmentation model ensured that no extraneous colors were present, leading to better performance. However, for image diversity (FID), Pix2Pix achieved the best score, while CycleGAN and the generative segmentation model scored poorly. This finding indicates that although the overall generated images closely matched the ground truth, their feature space representation differed significantly. The FFID, SSIM, and MAE scores from the evaluation across the four different models discussed above are shown in Fig. 9. This comparison highlights the performance variations among the models, providing a quantitative assessment of their effectiveness on the basis of these metrics.

To further evaluate the generated results, space syntax analysis was conducted to assess the spatial functional connectivity of the generated layouts. The functional spaces were categorized into two groups on the basis of their public attributes: public spaces (e.g., foyer, corridors, and atriums) and private spaces (e.g., exhibition halls, temporary exhibition halls, lecture halls, and spaces with other functions). Connectivity scores for each category and method were visualized in separate charts, with the x-axis representing methods, colors indicating functions, and the y-axis showing connectivity values normalized to a range of 0–100 for comparative purposes (Fig. 10).

The results of the LoRA revealed the absence of temporary exhibition hall and foyer functions. However, these two functions are essentially variants of exhibitions and corridors. Their omission has minimal impact on the overall museum layout, as these distinctions are often difficult to clearly define in museum designs. The high score assigned to exhibition functions in the first set of functional charts aligns well with the requirements of museum layouts. However, the relatively high score of the other functions is less justifiable, as this category should ideally have the lowest score. The fact that other functions scored higher than the atrium did is particularly unreasonable.

With respect to the GAN-based methods, Pix2Pix and CycleGAN exhibit similarities; both achieve high scores for the exhibition function, which is a favorable outcome. However, in the Pix2Pix results, the lecture hall score is the second highest, which is less optimal than that of Cycle-GAN, where the temporary exhibition hall score is the second highest. This prioritization better aligns with the needs of a museum layout. Notably, Pix2Pix assigns the lowest score to the other functions, which is appropriate, whereas in CycleGAN, the scores of the other functions are higher than that of the lecture hall, which is suboptimal for museum design. Overall, despite some discrepancies, the results generated by these two methods generally follow logical trends. In addition, the generative segmentation model yields the most desirable results for the exhibition hall function, achieving significantly higher scores than any other function does. This outcome suggests that the exhibition spaces are more prominent and perceptible to visitors, enhancing their spatial experience. However, the remaining functions, including the atrium, corridor, and foyer, which should have scored higher, have lower-than-expected scores. While the general trend is reasonable, the lack of clear differentiation among these secondary functions is a limitation.

When the spatial syntax analysis results of the generated layouts are compared with the original data, the connectivity scores for the exhibition halls in the generated results are generally higher. This finding indicates that the exhibition spaces in the generated layouts are more open and allow for better visual and experiential accessibility, enhancing visitors' perception of these spaces during their museum tours. For public spaces such as corridors, the original data outperform the generated results, with the original data showing higher scores for the atrium, a trend not reflected in the generated layouts. For the other functions, which should ideally have the lowest scores, only the original data conform to this expectation, while the generated results often assign higher scores to this category, making the original data superior in this respect.

Overall, owing to its similarity to the original data, Pix2Pix produces results where the functional connectivity scores closely resemble the original scores. In contrast, the other models generate results that deviate from the original data but still generally assign the highest scores to exhibition halls. This consistency is evident across all the generated models. However, for atrium spaces, the generated results consistently score lower than the original data do, highlighting an area where the generative models underperform.

4.2 Qualitative results

For qualitative analysis, we conducted evaluations with 21 designers and engineers across four aspects, with the results detailed in Table 2. With respect to the spatial distribution of the primary functions represented by the exhibition hall, temporary exhibition hall, lecture hall, and spaces with other functions in the generated images, the LoRA model achieved the highest scores, whereas CycleGAN scored the lowest. This finding indicates that the LoRA model produces more rational functional distributions than the less coherent outputs of the CycleGAN do. In the evaluation of public space configurations represented by the corridor, atrium, and foyer, the segmentation model attained the highest scores, whereas CycleGAN ranked lowest. These findings suggest that the segmentation model effectively integrates functional space through public spaces, whereas CycleGAN generates disorganized public space layouts that violate fundamental architectural spatial logic. With respect to functional demarcation quality as defined by wall placements, LoRA again achieved the highest score, with the segmentation model receiving the lowest score. This implies that LoRA generates clearer functional boundaries, whereas the segmentation model produces overlapping or ambiguous interfunctional divisions. In the overall plan generation assessment, LoRA maintained the highest score, with CycleGAN still achieving the lowest score. These results collectively demonstrate that the outputs of LoRA better align with the professional expectations of designers and engineers regarding functional composition and spatial organization (Fig. 11).

For qualitative analysis, space syntax results for the generated layouts were visualized across methods. Despite the suboptimal layouts generated by the CycleGAN and the generative segmentation model, their space syntax results were generally decent. This suggests that although the wall generation results lacked regularity and universality, the spatial connectivity of the generated layouts remained reasonable. Conversely, Pix2Pix and LoRA showed strong overall connectivity, with minimal areas of poor connectivity. In contrast, CycleGAN and the generative segmentation model exhibited larger proportions of poorly connected areas, with the generative segmentation model showing significant polarization―regions with either very high or very low connectivity―leaving minimal transitional spaces. This disparity underscores the varying strengths and weaknesses of each method in generating museum layouts. Space syntax analysis of the original data in the dataset reveals that the distributions of high and low scores in visual field analysis are generally consistent. However, there are notable differences in the specific spatial locations where scores are high or low.

Overall, in the original data, spaces located around the periphery tend to achieve higher scores in spatial syntax analysis. In contrast, the results generated by the models show the reverse trend, with peripheral spaces generally receiving lower scores and central spaces exhibiting higher scores. This distinction highlights a fundamental difference in the spatial distribution patterns between the original dataset and the generated layouts (Fig. 12).

5 Discussion: Museum floor plan design

Artificial intelligence, as an emerging general-purpose technology, has gained prominence across various industries since AlphaGo's groundbreaking achievement (Silver et al., 2016). The pivotal challenge lies in transforming this general technology into a specialized solution for professional fields. This study addresses this challenge through two primary approaches: dataset curation and evaluation framework development.

From the dataset perspective, AI technology operates by learning and extracting features from datasets and subsequently reconstructing these features under specific conditions to generate data resembling the original dataset (Krizhevsky et al., 2017). Our dataset exclusively comprises constructed museum floor plans, which inherently encapsulate the accumulated expertise and deliberate considerations of diverse designers. When a dataset embodies such professional architectural characteristics, trained models acquire foundational capabilities for serving the specialized domain of architectural floor planning.

With respect to the evaluation framework, this research supplements conventional AI performance metrics (e.g., comparisons with test data) with two additional dimensions: expert grading and space syntax analysis. Expert grading methodology is widely adopted in the architectural design literature (Fu et al., 2023; Min et al., 2023). Particularly for this study involving multiple generative methods, this approach effectively captures domainspecific professional evaluations.

Space syntax constitutes another critical dimension within our evaluation framework that provides an analytical perspective specifically tailored for museum-type buildings with distinct functional requirements (Hillier and Tzortzi, 2006). Conventional space syntax analysis typically employs Depthmap software, which requires DXF-formatted vector files as input, where walls possess mathematical expressions for efficient occlusion computation. However, most architectural floor plan generation models produce raster images, which typically require manual tracing in AutoCAD software to obtain vector graphs for analysis in Depthmap. This process is not only inefficient but may also lead to the loss of certain types of planar information during tracing, making it difficult to adapt to the analysis of large-scale AI-generated results. Consequently, this study implements space syntax analysis methodology on raster images and uses the ray casting method to address increased computational complexity. To verify the reliability of the proposed space syntax analysis method for raster images, this study selected a set of representative generated museum floor plans and compared the conventional Depthmap results with our raster-based outputs. The comparison results reveal overall consistency despite minor discrepancies. Specifically, the red areas on the heatmap of the analysis results indicate better spatial accessibility, whereas the blue areas indicate poorer accessibility. Although there are differences in the color gradation in local areas between the two methods, overall, the best and worst accessible spaces in the plan are consistent between the two methods (Fig. 13). These variations arise because the ray casting method simplifies computations differently than vector-based approaches do, Depthmap's rendering conventions differ from our method, and fundamental representational disparities exist between continuous pixel walls in raster images and discrete polylines in vectors. Manual vectorization cannot fully reconstruct raster wall data. Nevertheless, space syntax fundamentally evaluates holistic spatial relationships, and the consistent overall patterns validate our raster-based approach as being methodologically sound for this evaluation framework.

Furthermore, museums are complex systems in which various stakeholders have distinct functional requirements. Discussions on the generation of museum floor plans above focus on analyzing and evaluating generative models from a data-centric perspective. However, as a physical entity, a museum exerts real-world influence on citizens, making it essential to consider its architectural design comprehensively. In this section, a representative set of generated floor plans and their corresponding spatial syntax analysis results (Fig. 13) were selected for evaluation. These evaluations are conducted from three critical perspectives in museum architecture: designers, curators, and visitors.

Designers, typically comprising the primary architectural team responsible for the museum's design, prioritize comprehensive functionality and rational, efficient spatial arrangements that ensure seamless daily operations. Curators, often the initiators of exhibitions or events, require temporary and permanent exhibition spaces that cater to their specific needs for exhibition layouts. Visitors, including tourists and local residents, focus on their experiential perceptions of the museum, where their ability to observe and sense the environment is paramount. Spatial syntax analysis, particularly through visibility graph analysis, simulates human visual perception to identify areas that are easily observed and perceived. Thus, visitor-oriented evaluations rely primarily on spatial syntax analysis, whereas assessments for designers and curators focus on the generated floor plans. The following analysis compares the results of different models under a uniform input boundary, highlighting their strengths and weaknesses in comparison with the original data.

From the perspective of designers, the completeness of functions and the overall layout balance are crucial. The results generated by the LoRA model show a comprehensive functional distribution, with the lobby positioned externally, aligning well with its intended role. The “other functions” are located in corners so the exhibition circulation is not disrupted, which enhances the overall flow. Thus, the LoRA model performs effectively in this respect. The Pix2Pix model, on the other hand, produces results that are structurally similar to the original data and consistent with its high FID score. However, unlike the original data, where “other functions” occupy peripheral corners, Pix2Pix replaces these areas with atriums, leaving most other exhibition spaces intact. While creating open spaces in corners can be a design strategy, this result is less practical than LoRA's output. The results of CycleGAN and the generative segmentation model differ significantly from those of the other two models and feature more curved and flexible layouts. Although both models achieve functional completeness, CycleGAN results in the isolation of parts of the exhibition space with “other functions,” which disrupts circulation. The generative segmentation model generally provides smoother functionality and circulation but includes ambiguous small exhibition spaces. Overall, LoRA performs best in terms of functional distribution, followed by segmentation and CycleGAN, with Pix2Pix performing the worst.

From the curators' perspective, the layout of the exhibition spaces is paramount. The LoRA model generates relatively regular exhibition spaces, facilitating the easy organization of exhibitions. However, the small proportion of exhibition areas limits the scale and diversity of exhibition content. Pix2Pix, with results most similar to the original data, offers the largest and continuous exhibition spaces, which are ideal for hosting large-scale exhibitions. In contrast, the layouts of CycleGAN and the generative segmentation model deviate significantly from traditional exhibition layouts. While curved spaces are not uncommon in museums, they are typically custom designed on the basis of specific exhibition themes. The results of CycleGAN are asymmetric and feature small exhibition spaces, which are less favorable for curators. On the other hand, the generative segmentation model introduces notable variations from the original data while offering well-organized exhibition spaces, making it the most suitable for curators. Thus, segmentation ranks highest, followed by Pix2Pix, LoRA, and finally CycleGAN.

From the visitors' perspective, spatial experience is crucial, and spatial syntax analysis provides insights into how the generated spaces are perceived. For the LoRA model, corridor spaces score relatively high in visibility analysis, but only one exhibition hall achieves a high score, while the others score low, which is problematic. Furthermore, “other functions,” which should ideally remain inconspicuous to visitors, receive undesirably high scores, misaligning with typical visitor expectations in museums. Pix2Pix, owing to its structural similarity to the original data, generates spatial syntax results that closely mirror the original, with exhibition spaces scoring the highest, aligning well with visitor expectations. The segmentation and CycleGAN models, while unconventional, produce exhibition spaces with generally high syntax scores, “other functions” with the lowest scores, and corridors scoring in the middle range. These characteristics enhance visitor experience; segmentation and CycleGAN perform best in this respect, followed by Pix2Pix, with LoRA ranking last.

In summary, the LoRA model demonstrates the best performance from the perspective of designers and provides comprehensive functionality and a well-balanced layout. For curators, the generative segmentation model emerges as the most effective and offers exhibition spaces that align well with their requirements for organizing exhibitions. Similarly, from the perspective of visitors, the generative segmentation model performs the best, as its spatial configurations enhance the overall visitor experience on the basis of space syntax analysis. Considering all three perspectives, the generative segmentation model is deemed the most effective overall, followed by the LoRA model. Among the evaluated methods, the Pix2Pix model ranks third, whereas the CycleGAN model has the weakest performance (Table 3).

6 Conclusion

Architectural floor plan design constitutes a critical challenge in architectural practice. Automated plan generation can significantly reduce architects' workloads and enhance productivity. While substantial progress has been made in automated generation for residential domains, research on museum floor plan generation remains relatively limited. Current evaluations of automatically generated plans predominantly focus on model performance metrics, with few assessments from professional architectural perspectives. Space syntax emerges as a crucial analytical method that is particularly valuable for visitor-oriented buildings such as museums, and its application to evaluate AI-generated museum plans effectively supplements architectural evaluations of generated outcomes. In addition, expert grading could further evaluate the perceptual realism, functional adequacy, and human acceptance of AI-generated designs with specific professional domain knowledge.

In this study, we compiled and annotated a dataset of 263 museum floorplans using a color-coded system. We applied four methods across four distinct domains to train models on this custom dataset and evaluated the results from two perspectives. Specifically, the analysis of the results revealed that, first, for image structural information (SSIM), the GANs and LoRA methods achieved high scores, indicating that the spatial layout features of the generated images closely resembled those of the ground truth. Furthermore, compared with the other three models (CycleGAN, the segmentation model, and LoRA), Pix2Pix led in terms of the SSIM and the diversity metric FID but scored lowest in terms of the image detail reconstruction metric MAE. Second, space syntax analysis revealed that the connectivity of the exhibition space generated by the four models is greater than that of the original image, while the connectivity of the generated public spaces, such as the atrium, is lower than that of the original image. Third, the expert grading results indicate that the LoRA outputs better align with the professional expectations of designers and engineers regarding functional composition and spatial organization. With respect to the generative outcomes of different models from the perspectives of three key stakeholders (designers, curators, and visitors) in museum design, the generative segmentation model produces the most effective results, followed by the LoRA model, with Pix2Pix and CycleGAN ranking third and fourth, respectively. This evaluation represents one of the primary contributions of this study and provides a comprehensive framework for analyzing generative models in the context of museum-specific architectural needs.

Moreover, although space syntax is a well-established spatial analysis algorithm, its application to pixel-based generative floor plans remains unexplored. This study establishes a novel linkage between space syntax and automatically generated plans. On the one hand, this approach proposes a novel evaluation perspective based on spatial configuration for assessing AI-generated architectural floor plans that complements traditional image quality metrics such as the SSIM. On the other hand, pixel-based space syntax analysis significantly increases the efficiency of evaluating AI-generated architectural floor plan images.

Despite these contributions, several limitations exist in this study. To begin with, the dataset comprises only 263 samples, which is insufficient compared with residential datasets, which typically have more than 1000 samples. The inherent complexity of museum layouts also requires simplifications to ensure clarity for generative modeling, resulting in a dataset that may lack a comprehensive representation of intricate museum layouts. Furthermore, while this study focused on connectivity as a measure of spatial relationships, spatial syntax encompasses additional dimensions such as integration, which is critical for understanding potential spatial interactions. The omission of integration analysis is a notable limitation. Finally, this study establishes a comprehensive evaluation framework to filter models suitable for museum floor plan generation tasks and explores the applicability of different models to stakeholders. However, this study does not further optimize the AI design algorithm to directly improve the quality of the generated results, and how to integrate museum floor plan generation evaluation indicators into the generation model to optimize the floor plan generation results should be further explored in future research.

Overall, this research provides foundational work in generative museum floor plan design and contributes to dataset development, generative modeling, space syntax analysis and evaluations from the museum design perspective. Although the AI-driven generative model significantly increases the efficiency and analytical precision of museum graphic design, the graphic layout of museums, as a storehouse of collective memory and a carrier of cultural communication, should not only consider the spatial configuration itself but also further expand the synergistic fusion of the wisdom of multiple stakeholders; therefore, in future research, the intelligent generative model should be further explored with designers, curators, community representatives and end-users to shape a museum space with cultural significance and emotional resonance.

References

[1] Appel, A., 1968. Some techniques for shading machine renderings of solids. In: Proceedings of the April 30–May 2, 1968, Spring Joint Computer Conference, AFIPS '68 (Spring). Association for Computing Machinery, New York, NY, USA, pp. 37–45.

[2] Chen, R., Zhao, J., Yao, X., Jiang, S., He, Y., Bao, B., Luo, X., Xu, S., Wang, C., 2023. Generative design of outdoor green spaces based on generative adversarial networks. Buildings 13, 1083.

[3] Cui, X., Feng, X., Sun, S., 2024. Learning to generate urban design images from the conditional latent diffusion model. IEEE Access 12, 89135–89143.

[4] Doğan, N.B., Ayhan, B.U., Kazar, G., Saygili, M., Ayözen, Y.E., Tokdemir, O.B., 2022. Predicting the cost outcome of construction quality problems using case-based reasoning (CBR). Buildings 12, 1946.

[5] Eilouti, B., 2019. Shape grammars as a reverse engineering method for the morphogenesis of architectural façade design. Front. Archit. Res. 8, 191–200.

[6] Fu, B., Gao, Y., Wang, W., 2023. Dual generative adversarial networks for automated component layout design of steel frame-brace structures. Autom. ConStruct. 146, 104661.

[7] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y., 2020. Generative adversarial networks. Commun. ACM 63, 139–144.

[8] Güzelci, O.Z., Alaçam, S., Bekiroğlu, B., Karadag, I., 2024. A machine learning-based prediction model for architectural heritage: the case of domed Sinan mosques. Digit. Appl. Archaeol. Cult. Herit. 35, e00370.

[9] Han, Z., Chen, Y., 2024. Automatic generation of standard nursing unit floor plan in general hospital based on stable diffusion. Buildings 14, 2601.

[10] He, Q., Li, Z., Gao, W., Chen, H., Wu, X., Cheng, X., Lin, B., 2021. Predictive models for daylight performance of general floor-plans based on CNN and GAN: a proof-of-concept study. Build. Environ. 206, 108346.

[11] Hecht, M., Knutson, K., Crowley, K., Lyon, M., McShea, P., Giarratani, L., 2020. `How could the dinosaurs be so close to the future?': how natural history museum educators tackle deep time. Curator. Mus. J. 63, 39–54.

[12] Hegazi, Y.S., Tahoon, D., Abdel-Fattah, N.A., El-Alfi, M.F., 2022. Socio-spatial vulnerability assessment of heritage buildings through using space syntax. Heliyon 8, e09133.

[13] Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S., 2017. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17. Curran Associates Inc., Red Hook, NY, USA, pp. 6629–6640.

[14] Hillier, B., 1999. Space is the Machine: a Configurational Theory of Architecture. Cambridge University Press, Cambridge.

[15] Hillier, B., Hanson, J., 1984. The Social Logic of Space. Cambridge University Press, Cambridge.

[16] Hillier, B., Tzortzi, K., 2006. Space syntax: the language of museum space. In: Macdonald, S. (Ed.), A Companion to Museum Studies. John Wiley & Sons, Ltd, Hoboken, pp. 282–301.

[17] Ho, J., Jain, A., Abbeel, P., 2020. Denoising diffusion probabilistic models. In: Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20. Curran Associates Inc., Red Hook, NY, USA, pp. 6840–6851.

[18] Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., 2021. LoRA: Low-Rank Adaptation of Large Language Models.

[19] Isola, P., Zhu, J.-Y., Zhou, T., Efros, A.A., 2017. Image-To-Image Translation with Conditional Adversarial Networks. Presented at the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, pp. 5967–5976.

[20] Kahng, A.B., 2000. Classical floorplanning harmful?. In: Proceedings of the 2000 International Symposium on Physical Design, ISPD '00. Association for Computing Machinery, New York, NY, USA, pp. 207–213.

[21] Kajiya, J.T., 1986. The rendering equation. Siggraph Comput. Graph. 20, 143–150.

[22] Karadag, I., Güzelci, O.Z., Alaçam, S., 2022. EDU-AI: a twofold machine learning model to support classroom layout generation. Constr. Innov. 23, 898–914.

[23] Karimi, K., 2018. Space syntax: consolidation and transformation of an urban research field. J. Urban Des. 23, 1–4.

[24] Keshavarzi, M., Rahmani-Asl, M., 2021. GenFloor: interactive generative space layout system via encoded tree graphs. Front. Archit. Res. 10, 771–786.

[25] Kim, H., 2021. Evaluation of deep learning-based automatic floor plan analysis technology: an AHP-based assessment. Appl. Sci. 11, 4727.

[26] Ko, J., Ennemoser, B., Yoo, W., Yan, W., Clayton, M.J., 2023. Architectural spatial layout planning using artificial intelligence. Autom. ConStruct. 154, 105019.

[27] Kösenciğ, K.Ö., Okuyucu, E.B., Balaban, Ö., 2024. Structural plan schema generation through generative adversarial networks. Nexus Netw. J. 26, 409–427.

[28] Krizhevsky, A., Sutskever, I., Hinton, G.E., 2017. ImageNet classification with deep convolutional neural networks. Commun. ACM 60, 84–90.

[29] Li, Y., Xu, W., 2022. Using CycleGAN to achieve the sketch recognition process of sketch-based modeling. In: Yuan, P.F., Chai, H., Yan, C., Leach, N. (Eds.), Proceedings of the 2021 DigitalFUTURES. Springer, Singapore, pp. 26–34.

[30] Liu, J., Qiu, Z., Wang, L., Liu, P., Cheng, G., Chen, Y., 2024. Intelligent floor plan design of modular high-rise residential building based on graph-constrained generative adversarial networks. Autom. ConStruct. 159, 105264.

[31] Lu, S., 2010. Hidden orders in Chinese gardens: irregular fractal structure and its generative rules. Environ. Plann. Plann. Des. 37, 1076–1094.

[32] Merrell, P., Schkufza, E., Koltun, V., 2010. Computer-generated residential building layouts. ACM Trans. Graph. 29 (181), 1–181, 12.

[33] Min, X., Zheng, L., Chen, Y., 2023. The floor plan design method of exhibition halls in CGAN-assisted museum architecture. Buildings 13, 756.

[34] Mostafavi, F., Tahsildoost, M., Zomorodian, Z.S., Shahrestani, S.S., 2022. An interactive assessment framework for residential space layouts using pix2pix predictive model at the early-stage building design. Smart Sustain. Built. 13, 809–827.

[35] Mostafavi, F., van Engelenburg, C., Khademi, S., Vrachliotis, G., 2025. Floor plan generation: the interplay among data, machine, and designer. Int. J. Architect. Comput. 23, 370–386.

[36] Nourkojouri, H., Nikkhah Dehnavi, A., Bahadori, S., Tahsildoost, M., 2023. Early design stage evaluation of architectural factors in fire emergency evacuation of the buildings using Pix2Pix and explainable XGBoost model. J. Build Perform Simu. 16, 415–433.

[37] Pizarro, P.N., Hitschfeld, N., Sipiran, I., 2023. Large-scale multi-unit floor plan dataset for architectural plan analysis and recognition. Autom. ConStruct. 156, 105132.

[38] Qian, W., Xu, Y., Li, H., 2022. A self-sparse generative adversarial network for autonomous early-stage design of architectural sketches. Comput.-Aided CIV Inf. 37, 612–628.

[39] Rahbar, M., Mahdavinejad, M., Markazi, A.H.D., Bemanian, M., 2022. Architectural layout design through deep learning and agent-based modeling: a hybrid approach. J. Build. Eng. 47, 103822.

[40] Rani, S., Jining, D., Shah, D., Xaba, S., Singh, P.R., 2023. Exploring the potential of artificial intelligence and computing technologies in art museums. ITM Web Conf. 53, 01004.

[41] Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B., 2022. High-Resolution Image Synthesis with Latent Diffusion Models. Presented at the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, pp. 10674–10685.

[42] Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., Hassabis, D., 2016. Mastering the game of Go with deep neural networks and tree search. Nature 529, 484-+.

[43] Song, Z., 2025. Museum-Floor-Plan.

[44] Stiny, G., 1980. Introduction to shape and shape grammars. Environ. Plan B-Urban 7, 343–351.

[45] Stitt, F.A., 1984. Systems Graphics: Breakthroughs in Drawing Production and Project Management for Architects, Designers, and Engineers. McGraw-Hill, New York.

[46] Su, P., Lu, W., Chen, J., and, S.H., 2024. Floor plan graph learning for generative design of residential buildings: a discrete denoising diffusion model. Build. Res. Inf. 52, 627–643.

[47] Tang, Q., Zheng, L., Chen, Y., Yan, L., Chen, J., 2024. Artificial intelligence empowering museum space layout design: insights from China. PLoS One 19, e0310594.

[48] Turner, A., Doxa, M., O'Sullivan, D., Penn, A., 2001. From isovists to visibility graphs: a methodology for the analysis of architectural space. Environ. Plann. B 28, 103–121.

[49] van Aalst, I., Boogaarts, I., 2002. From museum to mass entertainment - the evolution of the role of museums in cities. Eur. Urban Reg. Stud. 9, 195–209.

[50] Wan, D., Zhao, X., Lu, W., Li, P., Shi, X., Fukuda, H., 2022. A deep learning approach toward energy-effective residential building floor plan generation. Sustainability 14, 8074.

[51] Wang, B., Faure, F., Pai, D.K., 2012. Adaptive image-based intersection volume. ACM Trans. Graph. 31 (97), 1–97, 9.

[52] Wang, T.-K., Duan, W., 2023. Generative design of floor plans of multi-unit residential buildings based on consumer satisfaction and energy performance. Dev. Built Environ. 16, 100238.

[53] Wang, X.-Y., Zhang, K., 2020. Generating layout designs from high-level specifications. Autom. ConStruct. 119, 103288.

[54] Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P., 2004. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Proc. 13, 600–612.

[55] Weber, R.E., Mueller, C., Reinhart, C., 2022. Automated floorplan generation in architectural design: a review of methods and applications. Autom. ConStruct. 140, 104385.

[56] Wineman, J.D., Peponis, J., 2010. Constructing spatial meaning spatial affordances in museum design. Environ. Behav. 42, 86–109.

[57] Wong, J.F., 2014. Conceptual engawa: the experience of ring-based circulation in Tadao Ando museums. Environ. Plan B-Urban 41, 229–250.

[58] Xu, S., Zhang, J., Li, Y., 2024. Knowledge-driven and diffusion model-based methods for generating historical building facades: a case study of traditional minnan residences in China. Information 15, 344.

[59] Yang, B., Jiang, T., Wu, W., Zhou, Y., Dai, L., 2022. Automated semantics and topology representation of residential-building space using floor-plan raster maps. IEEE J. Sel. Top. Appl. Earth Obs. Rem. Sens. 15, 7809–7825.

[60] Zeng, P., Gao, W., Yin, J., Xu, P., Lu, S., 2024. Residential floor plans: multi-conditional automatic generation using diffusion models. Autom. ConStruct. 162, 105374.

[61] Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O., 2018. The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Salt Lake City, pp. 586–595.

[62] Zhao, C.-W., Yang, J., Li, J., 2021. Generation of hospital emergency department layouts based on generative adversarial networks. J. Build. Eng. 43, 102539.

[63] Zhou, Y., Leng, H., Meng, S., Wu, H., Zhang, Z., 2024. StructDiffusion: End-to-end intelligent shear wall structure layout generation and analysis using diffusion model. Eng. Struct. 309, 118068.

[64] Zhu, J.-Y., Park, T., Isola, P., Efros, A.A., 2017. Unpaired image-to-image Translation Using cycle-consistent Adversarial Networks. Presented at the Proceedings of the IEEE International Conference on Computer Vision, pp. 2223–2232.

RIGHTS & PERMISSIONS

2095-2635/2025 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd.